A 2-CATEGORIES COMPANION
STEPHEN LACK
Abstract. This paper is a rather informal guide to some of the basic theory of 2-categories and bicategories, including notions of limit and colimit,
2-dimensional universal algebra, formal category theory, and nerves.
1. Overview and basic examples
1.1. The key players. There are bicategories, 2-categories, and Cat-categories.
The latter two are exactly the same (except that strictly speaking a Cat-category
should have small hom-categories, but that need not concern us here). The first
two are nominally different — the 2-categories are the strict bicategories, and not
every bicategory is strict — but every bicategory is biequivalent to a strict one,
and biequivalence is the right general notion of equivalence for bicategories and
for 2-categories. Nonetheless, the theories of bicategories, 2-categories, and Catcategories have rather different flavours.
An enriched category is a category in which the hom-functors take their values
not in Set, but in some other category V . The theory of enriched categories is now
very well developed, and Cat-category theory is the special case where V = Cat. In
Cat-category theory one deals with higher-dimensional versions of the usual notions
of functor, limit, monad, and so on, without any “weakening”. The passage from
category theory to Cat-category theory is well understood; unfortunately Catcategory theory is generally not what one wants to do — it is too strict, and fails
to deal with the notions that arise in practice.
In bicategory theory all of these notions are weakened. One never says that
arrows are equal, only isomorphic, or even sometimes only that there is a comparison 2-cell between them. If one wishes to generalize a result about categories to
bicategories, it is generally clear in principle what should be done, but the details
can be technically very difficult.
2-category theory is a “middle way” between Cat-category theory and bicategory theory. It uses enriched category theory, but not in the simple minded way
of Cat-category theory; and it cuts through some of the technical nightmares of
bicategories. The prefix “2-”, as in 2-functor or 2-limit, will always denote the strict
notion; although often we will use it to describe or analyze non-strict phenomena.
There are also various other related notions, which will be less important in
this companion. SSet-categories are categories enriched in simplicial sets; every 2category induces an SSet-category, by taking nerves of the hom-categories. Double
categories are internal categories in Cat. Once again every 2-category can be seen
as a double category. A slight generalization of double category allows bicategories
The support of the Australian Research Council and DETYA is gratefully acknowledged.
1
2
STEPHEN LACK
to fit into this picture. Finally there are the internal categories in SSet; both
SSet-categories and double categories can be seen as special cases of these.
1.2. Nomenclature and symbols. In keeping with our general policy, the word
2-functor is understood in the strict sense: a 2-functor between 2-categories A and
B assigns objects to objects, morphisms to morphisms, and 2-cells to 2-cells, preserving all of the 2-category structure strictly. We shall of course want to consider
more general types of morphism between 2-categories later on.
If “widget” is the name of some particular categorical structure, then there are
various systems of nomenclature for weak 2-widgets. Typically one speaks of pseudo
widgets for the up-to-isomorphism notion, lax widgets for the up-to-not-necessarilyinvertible comparison notion, and when the direction of the comparison is reversed,
either oplax widget or colax widget, depending on the specific case. But there are
also other conventions. In contexts where the pseudo notion is most important,
this is called simply a widget, and then one speaks explicitly of strict widgets in
the strict case. In contexts where the lax notion is most important (such as with
monoidal functors), it is this which has no prefix; and one has strict widgets in the
strict case or strong widgets in the pseudo.
As we move up to 2-categories and higher categories, there are various notions
of sameness, having the following symbols:
• = is equality
• ∼
= is isomorphism (gf = 1, f g = 1)
• ' is equivalence (gf ∼
= 1, f g ∼
= 1)
• ∼ is sometimes used for biequivalence, although it’s a bit unsatisfactory,
and it’s not at all clear what to do next.
In Sections 1.4 and 1.5 we look at various examples of 2-categories and bicategories. The separation between the 2-category examples and and the bicategory
examples is not really about strictness but about the sort of morphisms involved.
The 2-category examples involve functions or functors of some sort; the bicategory
examples (except the case of a monoidal category) involve more general types of
morphism such as relations. These “non-functional” morphisms are often depicted
p ) rather than an ordinary one (→). Typically the funcusing a slashed arrow ( →
tional morphisms can be seen as a special case of the non-functional ones. The
other special type of arrow often used is a “wobbly” one ( ); the denotes a weak
(pseudo, lax, etc.) morphism.
1.3. Contents. In the remainder of this section we look at examples of 2-categories
and bicategories. In Section 2 we begin the study of formal category theory, including some adjunction, extensions, and monads, but stopping short of the full-blown
formal theory of monads. In Section 3 we look at various types of morphism between
bicategories or 2-categories: strict, pseudo, lax, partial; and see how these can be
used to describe enriched and indexed categories. In Section 4 we begin the study
of 2-dimensional universal algebra, with the basic definitions and the construction
of weak morphism classifiers. This is continued in Section 5 on presentations for
2-monads, which demonstrates how various categorical structures can be described
using 2-monads. Section 6 looks at various 2-categorical and bicategorical notions
of limit and considers their existence in the 2-categories of algebras for 2-monads.
Section 7 is about aspects of Quillen model structures related to 2-categories and to
2-monads. In Section 8 we return to the formal theory of monads, applying some of
A 2-CATEGORIES COMPANION
3
the earlier material on limits. Section 9 looks at the formal theory of pseudomonads,
developed in a Gray-category. Section 10 looks at notions of nerve for bicategories.
There are relatively few references throughout the text, but Section 11 discusses
the main references and gives suggestions for further reading.
1.4. Examples of 2-categories. Cat is the mother of all 2-categories, just as
Set is the mother of all categories. From many points of view, it has all the best
properties as a 2-category (but not as a category: for example colimits in Cat are
not stable under pullback).
A small category involves a set of objects and a set of arrows, and also hom-sets
between any two objects. One can generalize the notion of category in various ways
by replacing various of these sets by objects of some other category.
(a) If V is a monoidal category one can consider the 2-category V -Cat of categories enriched in V ; these have V -valued hom-objects rather than hom-sets.
The theory works best when V is symmetric monoidal closed, complete, and
cocomplete. As for examples of enriched categories, one has ordinary categories
(V = Set), additive categories (V = Ab), 2-categories (V = Cat), preorders
(V = 2, the “arrow category”), simplicially enriched categories (V = SSet),
and DG-categories (V the category of chain complexes).
(b) More generally still, one can consider a bicategory W as a many-object version
of a monoidal category; there is a corresponding notion of W -enriched category:
see [3] or Section 3.1. Sheaves on a site can be described as W -categories for a
suitable choice of W .
(c) If E is a category with finite limits, one can consider the 2-category Cat(E)
of categories internal to E; these have an E-object of objects and an E-object
of morphisms. The theory works better the better the category E; the cases
of a topos or an abelian category are particularly nice. This includes ordinary
categories (E = Set), double categories (E = Cat), morphisms of abelian groups
(E = Ab), and crossed modules (E = Grp).
There is another class of examples, in which the objects are “categories with
structure”. The structure could be something like
(d)
(e)
(f)
(g)
(h)
category with finite products
category with finite limits
monoidal category
topos
category with finite products and coproducts and a distributive law
For most of these there are also analogues involving enriched or internal categories
with the relevant structure.
In each case you need to decide which morphisms to use. Normally you don’t
want the strictly algebraic ones (preserving the structure on the nose). Although
they can be technically useful, they are rare in nature. More common are the
“pseudo” morphisms: these are functors preserving the structure “up to (suitably
coherent) isomorphism”. In (e), for example, this would correspond to the usual
notion of finite-limit-preserving functor.
Sometimes, however, it’s good to consider an even weaker notion of morphism, as
in the 2-category MonCat of monoidal categories, monoidal functors, and monoidal
natural transformations. Here monoidal functors are the “lax” notion, involving
4
STEPHEN LACK
maps F A ⊗ F B → F (A ⊗ B), coherent, but not necessarily invertible. Here are
some reasons you might like this level of generality:
• Consider the monoidal categories Ab of abelian groups, with the usual
tensor product, and Set of sets, with the cartesian product. The forgetful
functor U from Ab to Set definitely does not preserve this structure, but
we have the universal bilinear map U G × U H → U (G ⊗ H), and this makes
U into a monoidal functor.
• A monoidal functor V → W sends monoids in V to monoids in W
• Suppose V and W are monoidal categories and F : V → W is a left adjoint
which does preserve the monoidal structure up to coherent isomorphism.
There is no reason why the right adjoint U should do so, but there will be
induced comparison maps U A ⊗ U B → U (A ⊗ B) making U a monoidal
functor. (Think of the tensor product as a type of colimit, so the left adjoint
preserves it, but the right adjoint doesn’t necessarily.) In fact the monoidal
functor U : Ab → Set arises in this way.
In fact the case of monoidal categories is typical. Given an adjunction F a U
between categories A and B with algebraic structure, to make the right adjoint
U a colax morphism is equivalent to making the left adjoint F lax, while if the
whole adjunction lives within the world of lax morphisms, then F is not just lax
but pseudo. This situation is called doctrinal adjunction [17].
For a further example, consider the structure of categories with finite coproducts.
For a functor F : A → B between categories with finite coproducts there are
canonical comparison maps F A + F B → F (A + B), and these make every such
functor uniquely into a lax morphism; it is a pseudo morphism exactly when it
preserves the coproducts in the usual sense. Thus in this case every adjunction
between categories with finite coproducts lives in the lax world, and the fact that
the left adjoint is actually pseudo reduces to the well known fact that left adjoints
preserve coproducts.
In the case of categories with finite products or finite limits, however, the lax
morphisms are the same as the pseudo morphisms; they are just the functors preserving the products or limits in the usual sense.
1.5. Examples of bicategories. Any monoidal category V determines a oneobject bicategory ΣV whose morphisms are the objects of V , and whose 2-cells are
the morphisms of V . The tensor product of V is the (horizontal) composition in
ΣV .
(i) Rel consists of sets and relations. The objects are sets and the morphisms
p Y are the relations from X to Y ; that is, the monomorphisms R X ×Y .
X→
This bicategory is ‘locally posetal’, in the sense that for any two parallel 1-cells,
there is at most one 2-cell between them. There is a 2-cell form R to S if and
only if R is contained in S as a subobject of X × Y ; in other words, if there is
a morphism R → S making the triangles in
R JJ
JJ
tt
tt
JJ
t
yt
%
X eJJ
:Y
t
t
JJ
JJ ttt
t
S
A 2-CATEGORIES COMPANION
5
commute. As usual, xRy means that (x, y) ∈ R. The composite of R X → Y
and S Y → Z is the relation R ◦ S defined by
x(R ◦ S)z ⇐⇒ (∃y)xRySz.
We get a 2-category biequivalent to this one by identifying isomorphic 1-cells;
this works for any locally posetal 2-category.
(j) Par consists of sets and partial functions. A partial function from X to Y is
a diagram X D → Y in Set; 2-cells and composition are defined as in Rel.
Again, we get a biequivalent 2-category by identifying isomorphic 1-cells.
(k) Span consists of sets and “spans” X ← E → Y in Set, with composition by
pullback, and with 2-cells given by diagrams such as
t E JJJ
JJ
tt
t
J%
ytt
X eJJ
9Y
t
t
JJ
JJ ttt
t
F
Unlike the previous two bicategories, this one is no longer locally posetal, so to
get a biequivalent 2-category we need to do more than just identify isomorphic
1-cells. There are general results asserting that any bicategory is biequivalent to
a 2-category, but in fact naturally occurring bicategories tend to be biequivalent
to naturally occurring 2-categories. In this case, we can take the 2-category
whose objects are sets and whose morphisms are the left adjoints Set/X →
Set/Y . Here the span
Xo
u
E
v
/Y
is represented by the left adjoint
Set/X
u∗
/ Set/E
v!
/ Set/Y
given by pulling back along u then composing with v.
(l) Mat has sets as objects, X × Y -indexed families (“matrices”) of sets as morphisms from X to Y . Composition is matrix multiplication, and 2-cells are
families of functions. This is biequivalent to Span, but we’ll see below that
spans and matrices become different when we start to consider enrichment
and internalization. A biequivalent 2-category consists of sets and left adjoints
SetX → SetY . (Here X ×Y → Set can be seen as a functor X → SetY , and so,
since SetX is the free cocompletion of X, as a left adjoint SetX → SetY .) This
is really just the same as the construction given for Span, since Set/X ' SetX ;
once again, though, when we start to enrich or internalize, the two pictures diverge.
p S, and
(m) Mod has rings as objects, left R-, right S-modules as 1-cells R →
p S and S →
p T is
homomorphisms as 2-cells. The composite of modules R →
given by tensoring over S. A biequivalent 2-category involves adjunctions
RMod SMod.
A ring is the same thing as an Ab-category (a category enriched in abelian
groups) with only one object. The underlying additive group of the ring is
the single hom-object; the multiplication of the ring is the composition. If we
identify rings with the corresponding one-object Ab-categories, then a module
p S becomes an Ab-functor R → [S op , Ab]
R→
6
STEPHEN LACK
But there is no reason to restrict ourselves to one-object categories, and
there is a bicategory Ab-Mod whose objects are Ab-categories, and whose
p B; that is, Ab-functors A → [B op , Ab].
1-cells are Ab-modules A →
More generally still, we can replace Ab by any monoidal category V with
coequalizers which are preserved by tensoring on either side, and there is then
a bicategory V -Mod of V -categories and V -modules: once again, if A and B
p B is a V -functor A → [B op , V ], or
are V -categories then a V -module A →
op
equivalently a left adjoint [A , V ] → [B op , V ], (and this last description gives
a 2-category).
There’s even, if you really want, a version with a bicategory W rather than
a monoidal category V .
Now let’s internalize and enrich the other examples.
(n) If E is a regular category, meaning that any morphism factorizes as a strong
epimorphism followed by a monomorphism, and the strong epimorphisms are
stable under pullback, then we can form Rel(E) whose objects are those of E
p Y are monomorphisms R X × Y . To compose
and whose morphisms X →
p Y and S : Y → Z we pullback over Y , but the resulting map into X ×Z
R:X→
need not be monic, so we need the factorization system to define composition.
It turns out that our assumption that strong epimorphisms are stable under
pullback is precisely what is needed for this composition to be associative.
(o) Similarly, if C is a category and M is a class of monomorphisms in C , then we
can look at Par(C , M), defined as above where the given monomorphism is in
M. There are conditions on M you need to make this work well: you want to
be able to pullback an M-map by an arbitrary map and obtain an M-map, and
you want M to be closed under composition and to contain the isomorphisms.
(p) If E has finite limits, we can look at Span(E) defined in an obvious way. You
need the pullbacks for composition to work. You don’t need any exactness
properties to get a category, but if you want to get a nice biequivalent 2category, you’ll need to start making more assumptions on E. It turns out that
Span(E) plays a crucial role in internal category: we shall see in Example 2.7
below an internal category in E is the same thing as a monad in Span(E).
(q) Mat, on the other hand, gets enriched rather than internalized. Then V -Mat
has sets as objects and V -valued matrices X × Y → V as morphisms. V Mat stands in exactly the same relationship to V -categories as Span(E) does
to categories in E . In the case V = Set of course V -Mat is just Mat, but
there is also another special case which we have already seen. Let V be the
arrow-category 2, consisting of two objects 0 and 1, and a single non-identity
arrow 0 → 1. This is cartesian closed (a V -category in this case is just a
preorder) and V -Mat in this case is Rel (we identify a subject of X × Y with
its characteristic function, seen as landing in 2).
1.6. Duality. A bicategory B has not one but three duals:
• B op is obtained by reversing the 1-cells
• B co is obtained by reversing the 2-cells
• B coop is obtained by reversing both
In the case of a monoidal category V , we can form the monoidal category V op by
reversing the sense of the morphisms; this reverses the 2-cells of the corresponding
A 2-CATEGORIES COMPANION
7
bicategory ΣV , so Σ(V op ) = (ΣV )co . Reversing the 1-cells of ΣV corresponds to
reversing the tensor of V , denoted V rev , so Σ(V rev ) = (ΣV )op .
2. Formal category theory
One point of view is that a 2-category is a generalized category (add 2-cells).
Another important one is that an object of a 2-category is a generalized category
(since Cat is the primordial 2-category). This is “formal category theory”: think
of a 2-category as a collection of category-like things.
You don’t capture all of V -category theory by thinking of V -categories as objects
of V -Cat, just as you don’t capture all of group theory by thinking of groups as
objects of Grp, but many things do work out well when we take this “elementfree” approach. In formal category theory you tend to avoid talking about objects
of a category, instead talking about morphisms (functors) into the category. Thus
morphisms become generalized objects (of their codomain) in exactly the same way
that morphisms in categories are generalized elements.
One of the starting points of formal category theory was Ross Street’s beautiful
work on the “formal theory of monads”. This was motivated by the desire to develop
a uniform approach to universal algebra for enriched and internal categories. It uses
all four dualities to incredible effect.
2.1. Adjunctions and equivalences. We start here with the notion of adjunction
in a 2-category (in other words, adjunction between objects of a 2-category — this
is not to be confused with adjunctions between 2-categories). In ordinary category
theory there are two main ways to say that a functor f : A → B is left adjoint
to u : B → A. First there is the local approach, consisting of a bijection between
hom-sets
B(f a, b) ∼
= A(a, ub)
for each object a ∈ A and b ∈ B, natural in both a and b. Alternatively, there is the
global approach, involving natural transformations η : 1A → uf and ε : f u → 1B
satisfying the usual triangle equations. Each can be generalized to the 2-categorical
setting.
Let K be a 2-category. Everything I’m going to say works for bicategories,
but let’s keep things simple; of course you can always replace a bicategory by a
biequivalent 2-category anyway.
An adjunction in K consists of 1-cells f : A → B and u : B → A, and 2-cells
η : 1A → uf and ε : f u → 1B satisfying the triangle equations. This is exactly
the global approach to ordinary adjunctions, with functors replaced by 1-cells, and
natural transformations by 2-cells. In a lot of 2-categories, this is a good thing to
study. We mentioned above the case MonCat. The study of adjunctions in Mod
is called Morita theory. In the case where η and ε are invertible, we have not just
an adjunction but an adjoint equivalence.
The local approach to adjunctions also works well here, provided that one uses
generalized objects rather than objects. For any 1-cells a : X → A and b : X → B,
there is a bijection between 2-cells f a → b and 2-cells a → ub. One now has
naturality with respect to both 1-cells x : Y → X, and 2-cells a → a0 or b → b0 .
This local-global correspondence can be proved more or less as in the usual case,
or it can be deduced from the usual case using a suitable version of the Yoneda
lemma. In fact the global-to-local part follows from the fact that 2-functors preserve
8
STEPHEN LACK
adjunctions, so that the representable 2-functors K (X, −) send the adjunction
f a u in K to an adjunction K (X, f ) a K (X, u) in Cat, between K (X, A)
and K (X, B), and so the usual properties of adjunctions give the correspondence
between f a = K (X, f )a → b and a → K (X, u)b = ub.
The contravariant representable functors
K (−, X) : K
op
→ Cat
also preserve adjunctions. This prepares you for:
Exercise 2.1. f is a left adjoint in K if and only if it is a right adjoint in K
and only if it is a right adjoint in K op .
co
if
Exercise 2.2. A morphism f : A → B in a 2-category K is said to be an equivalence if there exist a morphism g : B → A and isomorphisms gf ∼
= 1B .
= 1A and f g ∼
Show that for any equivalence f these data can be chosen so as to give an adjoint
equivalence. Hint: you can keep the same f and g; you’ll need to change at most
one of the isomorphisms.
Considering an adjunction f a u in K as an adjunction in K op , and using the
local approach, we see that to give a 2-cell s → tf is the same as to give a 2-cell
su → t. Even in the case K = Cat this is not as well known as it should be.
More generally, given a pair of adjunctions f a u and f 0 a u0 , we have bijections
between 2-cells f 0 a → bf , 2-cells a → u0 bf , and 2-cells af 0 → u0 b: squares
A
a
A0
f
⇒
f0
/B
b
/ B0
correspond to squares
Ao
a
A0 o
u
⇒
u0
B
b
B0
These pairs of 2-cells are called mates.
2.2. Extensions. Extensions generalize Kan extensions. They provide limit and
colimit notions for objects of a 2-category, generalizing the usual notions for categories.
Let K be a 2-category. What is the universal solution to extending f along j?
BO
j
A
KS f
/C
Such a universal solution is denoted lanj f ; by universal we mean that it induces a
bijection
f −→ gj
lanj f −→ g
A 2-CATEGORIES COMPANION
9
for any g : B → C. When such a lanj f exists in K , it is called a left extension of
f along j.
A colimit is absolute if it is preserved by any functor; similarly we say that
the left extension lanj f is absolute if composing with g : C → D gives another
extension, so that glanj f = lanj (gf ).
Consider the case K = Cat. There would be such a bijection if lanj f were
the left Kan extension Lanj f of f along j, as indeed the notation is supposed to
suggest. In the case of (pointwise) left Kan extensions, we have a coend formula
Z a
(lanj f )b =
B(ja, b) · f a.
Alternatively the right hand side can be expressed using colimits: given b we can
form the comma category j/b, with pairs (a ∈ A, ja → b) as objects, and the
canonical functor d : j/b → A, then the coend on the right hand side is (canonically
isomorphic to) the colimit of f d : j/b → C.
(Kan extensions which are not ‘pointwise’ — in other words, which don’t satisfy
this formula — can exist if C is not cocomplete, but should be regarded as somewhat
pathological.)
How might we express this formula so that it makes sense in an arbitrary 2category? Once again, the answer will involve generalized objects. Consider an
object b ∈ B as a morphism b : 1 → B, and then consider the diagram
1O
c
j/b
b
KS d
/B
O
j
/A
KS f
lanj f
/C
Ra
in which j/b is the comma category. The coend
B(ja, b) · f a is isomorphic to
the colimit of f d, as we saw, but the colimit of f d is itself isomorphic to the left
Kan extension of f d along the unique map j/b → 1. A careful calculation of the
isomorphisms involved reveals that the coend formula amounts to the assertion that
the diagram above is a left extension.
This motivates the definition of pointwise extension in a general 2-category K
with comma objects. We say that the left extension lanj f is pointwise if, for any
b : X → B, when we form the comma object the 2-cell
XO
c
j/b
b
KS d
/B
O
j
/A
KS f
lanj f
/C
exhibits (lanj f )b as lanc (f d).
This agrees with the usual definition in the case K = Cat, works perfectly in
the case of Cat(E), and captures many but not all features in V -Cat.
Let’s leave the pointwise aspect aside and go back to extensions.
• A left extension in K
• A left extension in K
• A left extension in K
co
(reverse the 2-cells) is called a right extension.
(reverse the 1-cells) is called a left lifting.
coop
(reverse both) is called a right lifting.
op
10
STEPHEN LACK
The right lifting r : X → A of b : X → B through f : A → B is characterized by
a bijection
f a −→ b
a −→ r
which is a sort of internal-hom; indeed, in the one-object case, where the composite
f a is given by tensoring, it really is an internal hom. Some people use the notation
r = f \ b for this lifting.
A special case is adjunctions. Given f a u : B → A, we have a bijection
f a −→ b
a −→ ub
and so ub = f \ b is the right lifting of b through f . In particular, u is the right
lifting of the identity 1B through f . Conversely, a right lifting u of the identity
through f is a right adjoint if and only if it is absolute; in other words, if ub is the
right lifting of b through f for all b : X → B; in symbols f \ b = (f \ 1)b.
Dually, given an adjunction f a u : B → A we have a bijection
xu −→ y
x −→ yf
and so yf = ranu y and f = ranu 1B ; while in general a right extension f = ranu 1B
of the identity is a left adjoint of u if and only if it is absolute.
A bicategory is said to be closed if it has right extensions and right liftings.
In the one-object case, this means that the endofunctors − ⊗ c and c ⊗ − of the
monoidal category have right adjoints for any object c.
We saw that pointwise left extensions in Cat are given by colimits. Thus the
existence of left extensions is some kind of internal cocompleteness condition. So in
2-categories like Cat(E) or V -Cat they will exist only in some cases. In bicategories
like V -Mod, on the other hand, all extensions exist (provided that V is itself
complete and cocomplete).
Let me point out a little lemma which everyone knows for Cat, but which is
true for 2-categories basically because everything is representable. A morphism
f : A → B in a 2-category K is said to be representably fully faithful if K (X, f ) :
K (X, A) → K (X, B) is a fully faithful functor for all objects X of K . For
K = Cat this is equivalent to f being fully faithful.
Lemma 2.3. Let f a u be an adjunction in a 2-category K for which the unit
η : 1 → uf is invertible. Then f is representably fully faithful.
Similarly, under the same hypotheses, u will be (representably) “co-fully-faithful”,
in the sense that each K (u, X) : K (B, X) → K (A, X) is fully faithful.
2.3. Monads. Just as in ordinary category theory, an adjunction f a u : B → A
in a 2-category induces a 2-cell t = uf , with 2-cells η : 1 → uf = t, given by the
unit of the adjunction, and a multiplication µ = uεf : t2 = uf uf → uf = t, where
ε : f u → 1 is the counit. This η and µ make t into a monoid in the monoidal
category K (A, A).
More generally, a monad in a 2-category K on an object A ∈ K consists of
a 1-cell t : A → A equipped with 2-cells η : 1 → t and µ : t2 → t satisfying the
usual (associative and identity) equations; the situation of the previous paragraph
is a special case. One often speaks simply of a monad (A, t), when η and µ are
understood.
A 2-CATEGORIES COMPANION
11
The case K = Cat is just the usual notion of monad on a category A. (This is
sometimes called a monad in A, but this is to be avoided: it is in K and on A.)
Example 2.4. Monads in Cat are the usual monads. Monads in V -Cat or Cat(E)
are the obvious notion of enriched or internal monad. Monads in MonCat are
called monoidal monads.
Example 2.5. Monads in the one-object 2-category ΣV are monoids in the strict
monoidal category V . Conversely, a monad in an arbitrary 2-category K , on an
object X of K , is a monoid in the (strict) monoidal category K (X, X). There are
analogous facts for bicategories and (not necessarily strict) monoidal categories.
p E0 , in the
Example 2.6. Monads in Rel. We have a set E0 ; a relation t : E0 →
form of a subset R of E0 × E0 ; the “identity” 1 → t amounts to the assertion that
the relation R is reflexive, and the multiplication to the fact that R is transitive.
The associative and unit laws are automatic.
p E0 ,
Example 2.7. Monads in Span(E). We have an object E0 , a 1-cell t : E0 →
as in
E1 K
KKKc
d sss
KK
ysss
%
E0
E0
(a directed graph in E), with a multiplication
µ : E1 ×E0 E1 → E1
from the object of composable pairs to the object of morphisms, giving a composite; associativity of the monad multiplication is precisely associativity of the
η
composition. Similarly the unit 1 −→ t gives E0 → E1 since the identity span is
E0
sss
sy ss
E0 K
KKK
KK
%
E0
and the unit laws for the monad are precisely the identity laws for the internal
category. Thus a monad in Span(E) is the same as an internal category in E .
This is one of the main reasons for considering the span construction.
Example 2.8. Monads in V -Mat. We have an object X, which is just a set, a 1p X, in the form of a matrix X ×X → V , which we think of as sending (x, y)
cell X →
to a hom-object C (x, y). The multiplication map goes from the matrix product, as
in
X
C (y, z) ⊗ C (x, y) −→ C (x, z)
y
and gives a composition map. Once again the associative and identity laws for the
composition are precisely the associative and unit laws for the monad, and we see
that a monad in V -Mat is the same as a category enriched in V .
In the special case V = 2 we have V -Mat = Rel, and so we recover the
observation, made in Example (q) above, that a category enriched in 2 is just a
preorder (a reflexive and transitive relation).
A morphism of monads from (A, t) to (B, s) consists of a 1-cell f : A → B
equipped with a 2-cell ϕ : sf → f t, satisfying two conditions: see [46] or Section 8
12
STEPHEN LACK
below. A morphism of monads in Span(E) is not an internal functor, since it
p F0 between the objects of objects, rather than
would involve a 1-cell (a span) E0 →
a morphism in E. In order to get internal functors, we need to consider not Span(E)
itself, but rather Span(E) equipped with the class of “special” 1-cells consisting of
those spans whose left leg is the identity; these can of course be identified with the
1-cells in E. An internal functor will turn out to be a monad morphism, for which
p F0 is “special”.
the span E0 →
The case of enriched functors is similar: one needs to keep track of the 1-cells in
V -Mat which are really just functions.
To get (enriched or internal) natural transformations, you do not use the obvious
notion of monad 2-cells as in [46], but rather those of [34]; once again see Section 8
below.
3. Morphisms between bicategories
3.1. Lax morphisms. We talked before about the virtues of monoidal functors
between monoidal categories. The corresponding morphisms between bicategories
are the lax functors (originally just called morphisms of bicategories by Bénabou).
A lax functor A → B sends objects A ∈ A to objects F A ∈ B, has functors
F : A (A, B) → B(F A, F B) (thus preserving 2-cell composition in a strict way),
and has comparison maps ϕ : F g · F f → F (gf ) and ϕ0 : 1F A → F (1A ) and some
coherence conditions, which are formally identical to those for monoidal functors.
All the good things that happen for monoidal functors happen for lax functors.
For example, monoidal functors take monoids to monoids, and lax functors take
monads to monads. (Recall that a monad in B on an object X is the same as a
monoid in the monoidal category B(X, X).)
As a very special case, consider the terminal 2-category 1. This has a unique
object ∗, and a unique monad on ∗ (the identity monad). Then for any lax functor
1 → B, the object ∗ gets sent to F ∗ = A, the identity 1 is sent to F 1 = t, the
comparison maps become µ : tt → t and η : 1 → t, and the coherence conditions
make this precisely a monad. In fact, monads in B are the same as lax functors
1 → B. For Bénabou, this was the main reason to consider lax morphisms of
bicategories, rather than the stronger version.
In particular, V -categories are the same as monads in V -Mat, and so also the
same as lax functors 1 → V -Mat. This is the same as a set X together with a lax
functor
Xch −→ ΣV
where Xch is X made into a chaotic bicategory (also called indiscrete: every homcategory Xch (x, y) is trivial). Why? We send each x to ∗, we have a functor
1 = Xch (x, y) → ΣV (∗, ∗) = V
picking out the hom-object C (x, y) ∈ V , and the lax comparison maps ϕ become
the composition and identity maps.
If we replace ΣV by an arbitrary bicategory W , we get the notion of a W -enriched
category: a set X with a lax functor
Xch −→ W
Another way to think about Xch , as a bicategory, is to say that the unique
map X → 1 is fully faithful. But we can also consider, more generally, a pair of
A 2-CATEGORIES COMPANION
13
bicategories with a partial map
A
Nn D
` `
}}
}
` `
}}
}
}~ }
B
where the hooked arrow ,→ denotes a fully faithful strict morphism, and the wobbly
map
denotes a lax functor. This partial map is called a 2-sided enrichment or
a category enriched from A to B . If A is 1, it’s just a category enriched over
B. Using the notion of composition for these things is very helpful in analyzing
the change of base between different bicategories. For example, a B-category is a
partial map from 1 to B; this can be composed with a partial map from B to C
to get a C -category.
3.2. Pseudofunctors and 2-functors. A pseudofunctor (or homomorphism of
bicategories) is lax functor for which ϕ and ϕ0 are invertible.
Example 3.1. For a bicategory B, the representables
B
B(B,−)
/ Cat
are pseudofunctors, not strict in general.
Example 3.2 (Indexed categories). A pseudofunctor B op → Cat is sometimes
called a B-indexed category. Often B itself will just be a category (no non-identity
2-cells), in which case such a pseudofunctor corresponds to a fibration E → B in
the Grothendieck picture.
An important property of pseudofunctors not shared by lax functors is that they
preserve adjunctions. Consider a pseudofunctor F : A → B, and an adjunction
f a u : B → A in A , with unit η : 1A → uf and ε : f u → 1B . We may apply F to
f and u to get F f : F A → F B and F u : F B → F A, and now the composite 2-cells
F f.F u
1F A
ϕ
/ F (f u)
Fε
/ F 1B
ϕ−1
0
/ 1F B
ϕ0
/ F 1A
Fη
/ F (uf )
ϕ−1
/ F u.F f
provide the unit and counit for an adjunction F f a F u. This fails for a general lax
functor F .
If ϕ and ϕ0 are not just invertible, but in fact identities, then one speaks of a
strict homomorphism; or, in the case of 2-categories, of a 2-functor. Note that in the
bicategory case the associativity and identity constraints must still be preserved:
this is the content of the coherence condition for ϕ and ϕ0 .
2-functors are much nicer to work with, but often it is the pseudofunctors which
arise in nature. One reason you might prefer 2-functors is so as not to have to worry
about coherence. Furthermore, 2-functors have better properties than pseudofunctors: for example, the category 2-Cat of 2-categories and 2-functors has limits and
colimits, but the category 2-Catps of 2-categories and pseudofunctors does not.
14
STEPHEN LACK
For example the diagram
7/ 1
ppp
p
p
ppp
ppp
p
p
p
1>
>>
>>
>>
/2
1
0
has no pushout: such a pushout would have to have morphisms 0 → 1 → 2 and a
composite, but in some other cocone we have no way to decide where to send the
composite. If, however, we made 2-Catps into a tricategory, then it would have
trilimits (the relevant “weak” notion of limits for tricategories).
On the other hand, even if you start in the world of 2-categories and 2-functors,
F
you may be forced out of it. A 2-functor A −→ B is a biequivalence if A (A, B) →
B(F A, F B) are equivalences and it is “bi-essentially surjective”, in the sense that
for all X ∈ B, there exists an A ∈ A and an equivalence F A ' X in B. This is
the “right notion” of equivalence for 2-functors.
The point is that you’d like something going back the other way from B to A .
Well you do have something, but it’s just not a 2-functor in general. Given X ∈ B,
x
pick A ∈ A and F A ' X and let GX = A. Given X −→ Y , we can bring it across
the equivalences F A ' X and F B ' Y to get x : F A → F B, and since F is locally
an equivalence, x ∼
= F a for some a : A → B; let Gx = a. This all works, but since
everything is only defined up to isomorphism, there’s no way you can possibly hope
for G to preserve things strictly.
There is a Quillen model structure on 2-Cat — see Section 7.5 below — for
which the weak equivalences are the biequivalences, and clearly getting a 2-functor
B → A is going to have something to do with B being cofibrant.
3.3. Higher structure. As well as lax (and other) morphisms between bicategories, there is higher structure. Given morphisms F, G : A → B, one can consider
families αA : F A → GA of morphisms in B indexed by the objects of A , and
subject to (lax, oplax, pseudo, or strict) naturality conditions. There is even a
further level of structure, consisting of morphisms between such transformations:
these are called modifications.
4. 2-dimensional universal algebra
There are various categorical approaches to universal algebra: theories, operads,
sketches, and others, but I’ll mostly talk about monads, although you may see
parallels with operads and with theories if you know about those.
The ordinary universal algebra picture you might have in mind is monoids (or
groups, rings, etc.) living over sets. But our algebras don’t have to be single-sorted;
they could live over some power of sets. Abstractly, of course, we could be living
over almost everything. A good many-sorted example to have in mind is the functor
category [C , Set] living over [obC , Set], for a small category C . If C has one object,
then we may identify C with the monoid M of its arrows, and the functor category
is then the category of M -sets.
A 2-CATEGORIES COMPANION
15
When we come to 2-categories, we might generalize monoids over sets to monoidal
categories over categories; or (also living over categories) categories with finite products, or with finite coproducts, or with both, or with finite products and finite
coproducts and a distributive law.
For an example of the many-sorted case, let B be a small bicategory. There is
a 2-category Hom(B, Cat) of homomorphisms from B to Cat (B-indexed categories), whose morphisms and 2-cells are the pseudonatural transformations and
modifications defined below.
Hom(B, Cat)
[obB, Cat]
is an example of the sort of algebraic structure we have in mind.
In the next two sections there is a lot of interplay between 2-category theory and
Cat-category theory. Since I don’t want to assume enriched category theory, I’ll
tend to describe the ordinary (unenriched) setting, take it for granted that one can
modify this to get a Cat-enriched version, and concentrate more on how to modify
this to do the proper 2-categorical one.
4.1. 2-monads. We continue to follow the convention that the prefix 2- indicates
a strict notion. Thus a 2-monad consists of a 2-category K equipped with a 2functor T : K → K , and 2-natural transformations m : T 2 → T and i : 1 → T ,
satisfying the usual equations for a monad. In other words, this is a monad in the
(large) 2-category of 2-categories, 2-functors, and 2-natural transformations. (This
could be made into a 3-category, but we don’t need to do so for this observation.)
There is a good theory of enriched monads — this was one of the motivations
of the formal theory of monads — and 2-monads are just V -monads in the case
V = Cat.
A (strict) T -algebra is the usual thing, an object A ∈ K with a morphism
a : T A → A satisfying the usual equations, written (A, a). Once again, this is the
strict (or Cat-enriched) notion.
Remark 4.1. There are pseudo and lax notions of monad and of algebra, but they
seem to be less important than the strict ones. The main reason for this is that
the actual structures one wants to describe using 2-dimensional monads are the
strict algebras for strict monads in a fairly straightforward way — an example is
given below — whereas identifying the structures of interest with pseudoalgebras is
rather more work. A secondary reason is that in reasonable cases a pseudomonad T
can be replaced by a strict monad T 0 whose strict algebras are the pseudoalgebras
of T .
It is when we come to the morphisms of algebras, however, that we are forced
to depart the strict setting. A lax T -morphism (A, a) → (B, b) is a morphism
f : A → B in K , equipped with a 2-cell
TA
a
A
Tf
f
f
/ TB
b
/B
16
STEPHEN LACK
satisfying two coherence conditions:
T 2A
T 2f
T f
Ta
TA
/ T 2B
Tf
T 2A
TA
=
f
A
f
i
TA
f
/B
A
f
f
mB
/ TB
f
b
/B
/B
iB
Tf
f
a
A
A
/B
/ T 2B
Tf
a
b
A
mA
Tb
/ TB
a
T 2f
f
/ TB
=
1
b
/B
A
1
f
/ B.
Note that the outer 1-cells are the same (I wouldn’t write this down if they weren’t),
and that empty regions commute, and are deemed to contain the relevant identity
2-cell.
P
Let’s do a baby example: K = Cat and T A = n An the usual free monoid
construction. The T -algebras are strict monoidal categories, and a lax morphism
involves 2-cells
P
P n
n
/
nB
nA
⊗
⊗
f
/B
A
so we have transformations
f (a1 ) ⊗ . . . ⊗ f (an ) −→ f (a1 ⊗ . . . ⊗ an ).
for each n. The definition of monoidal functor only mentions the cases n = 0 and
n = 2, but all the others can be built up from these in an obvious way; the coherence
conditions for lax T -morphisms say that you do build them up in this sensible way,
and that the coherence conditions for monoidal functors are satisfied.
So for this T , the lax morphisms are precisely the monoidal functors. This
provides a practical motivation for the definition of lax T -morphism. Here’s a
theoretical one. There’s a 2-category Lax(2, K ) where 2 is the arrow category. In
detail:
• An object is an arrow a : A0 → A in K
• A 1-cell is a square
A0
a
A
f0
f
/ B0
ϕ
b
/B
A 2-CATEGORIES COMPANION
17
• A 2-cell (f, ϕ, f 0 ) → (g, ψ, g 0 ) consists of 2-cells α : f → g and α0 : f 0 → g 0
making the diagram
bα0
bf 0
/ bg 0
ϕ
ψ
fa
αa
/ ga
commute.
Since this is functorial in K , the 2-monad T induces a 2-monad Lax(2, T ) on
Lax(2, K ). Then a (strict) Lax(2, T )-algebra is precisely a lax T -morphism. The
coherence conditions for lax morphisms become the usual axioms for algebras.
Similarly, a T -transformation between lax T -morphisms (f, f ), (g, g) : (A, a) →
(B, b) is a 2-cell ρ : f → g in K such that
T Aa
A
T ρ
g
g
)
f
5 TB = TA
a
b
A
4B
)
TB
f
ρ
b
& 8B
In the baby example, for n = 2 this says that
/ f (a1 ⊗ a2 )
f a1 ⊗ f a2
ρ
ρa1 ⊗ρa2
/ g(a1 ⊗ a2 )
ga1 ⊗ ga2
which is exactly the condition for ρ : f → g to be a monoidal natural transformation.
Exercise 4.2. Play the Lax(2, K ) game with T -transformations: find a 2-category
K 0 and a 2-monad T 0 on K 0 whose algebras are the 2-monads.
There is a 2-category T -Alg` of T -algebras, lax T -morphisms, and T -transformations,
and a forgetful 2-functor
U
`
T -Alg` −→
K
and in some cases, such as that of monoidal categories, this is the 2-category of
primary interest, but often the pseudo case is more important (and of course strong
monoidal functors are themselves important). If f is invertible, we say that (f, f )
is a pseudo T -morphism or just a T -morphism (privileging these over the strict or
the lax). These are the morphisms of the 2-category T -Alg of T -algebras, pseudo
T -morphisms, and T -transformations; it has a forgetful 2-functor
U
T -Alg −→ K .
When f is an identity we have a strict T -morphism. Of course this just means
that the square commutes, and we have a morphism in the usual unenriched sense,
but it is still useful to think of the identity 2-cell as being “an f ”, since it is
used in the condition on 2-cells. The T -algebras, strict T -morphisms, and T transformations form a 2-category T -Algs with a 2-functor
U
s
T -Algs −→
K.
18
STEPHEN LACK
Each of these 2-categories has the same objects, and we have
T -Algc
:
uu
u
u
J`
u
uu
uu
*
/ T -Alg
/ T -Alg`
T -Algs
JJ J
t
JJ
t
t
JJ
tt
JJ
t
J$ zttt
K
where T -Algc is the 2-category of colax morphisms, defined like lax morphisms
except that the direction of the 2-cell is reversed. I won’t worry too much about
them since they can be treated as the lax morphisms for an associated 2-monad on
K co .
At this point we need to start making some assumptions. To start with, suppose
that K is cocomplete, and that T has a rank, which means that T : K → K
preserves α-filtered colimits for some α. For ordinary monads on categories, it says
that we can describe the structure in terms of operations which may not be finitary,
but are at least α-ary for some regular cardinal α. The famous example of a monad
on Set which is not α-filtered for any α is the covariant power set monad.
Under these conditions
J
T -Algs −→ T -Alg
J
`
T -Algs −→
T -Alg`
have left adjoints. What does this mean? Among other things it means that for
each algebra A there is an algebra A0 and bijections
A
B
0
A →B
where the wobbly arrow denotes a weak morphism and the normal arrow a strict
one. Here “weak” might mean either pseudo or lax, depending on the context; of
course there will be a different A0 depending on whether we consider the pseudo or
the lax case.
These are 2-adjunctions, so these bijections are just part of isomorphisms of
categories
T -Algs (A0 , B) ∼
= T -Alg(A, JB)
2-natural in A and B. We usually omit writing the J, since it is the identity on
objects. We say that such an A0 classifies weak morphisms out of A. From this we
get a unit
p: A
A0
and counit
q : A0 → A
and one of the triangle equations tells you that qp = 1. An unfortunate consequence
of the (otherwise reasonable) notation A0 , is that the left adjoint to J : T -Algs →
T -Alg is sometimes saddled with the rather embarrassing name ( )0 ; I shall call it
Q instead.
4.2. Sketch proof of the existence of A0 .
A 2-CATEGORIES COMPANION
19
Step 1. T -Algs is cocomplete.
This part is entirely “strict”: it is really an enriched category phenomenon, and
not really any harder than the corresponding fact for ordinary categories. It is here
that you use the assumptions on K and T .
Colimits of algebras, as we know, are generally hard. The problem is essentially
that algebras are a “quadratic” notion, involving a : T A → A with two copies of
A. We “linearize” and it becomes easy. What does that mean?
Take the T -algebra (A, a : T A → A), forget the axioms, and also forget that the
two A’s are the same, so consider it only as a map a : T A → A1 . This defines the
objects of a new category, whose morphisms are squares of the form
TA
Tf
a
A1
/ TB
b
f1
/ B1 .
With the obvious notion of 2-cell this becomes a 2-category; in fact it is just the
comma 2-category T /K . The point is that we have a full embedding
T -Algs ,→ T /K
since by the unit condition for algebras, any morphism in T /K between algebras
must have f = f1 and so be a strict T -morphism. It is this T /K which is the
“linearization” of T -Algs , and colimits in it are easy. Say we have a diagram of
things T Ai → Bi . Take the colimits in K and take the pushout
colimT Ai
/ colimBi
T colimAi
/B
to get the colimits in T /K .
The hard bit, which I’ll leave out, is the construction of a reflection T /K →
T -Algs (a left adjoint to the inclusion). This is where we use the assumption on T .
There are some transfinite calculations, as you might expect given the condition on
α-filtered colimits.
Note, however, that should T preserve all colimits, then this Step 1 becomes
easy: the colimits are constructed pointwise. In particular, this is true in the case
of categories of diagrams (T -Algs = [B, Cat]).
In fact when we come to step 2, we’ll see that only certain (finite) colimits
in T -Algs are actually needed, and if T should preserve these colimits, as does
sometimes happen, then once again the proof simplifies.
Step 2. Let (A, a) be an algebra; we want to construct the pseudomorphism classifier A0 using colimits in T -Algs . A lax T -morphism (A, a) → (B, b) consists of
various data in K , and we want to translate all that data into T -Algs .
A lax T -morphism A → B consists of
• A morphism f : A → B in K , which becomes a morphism g : T A → B in
T -Algs , where g = b · T f .
20
STEPHEN LACK
• A 2-cell
/ TB
Tf
TA
a
A
f
b
/B
f
in K , which becomes a 2-cell
mA
T 2A
Ta
TA
ζ
g
/ TA
g
/B
in T -Algs , since
b.T (f a) = b.T f.T a = g.T a
b.T (b.T f ) = b.T b.T 2 f = · · · = g.mA.
• The condition f .iA = id corresponds to saying that ζ.T iA = id
• The other condition becomes
mA
TB 2 A8
888
mT A 8T8 a
88
3
TB A
T A:
::
::
:
mA
T 2 a ::
T 2A
Ta
/ TA
44
44
ζ
44g
44
g
/B
D
ζ
g
/ TA
=
TB 2 A
mT A T mA
T 3 A:
::
::
:
T 2 a ::
T 2A
/ TA
B 4
444
mA 44g
44
/ T 2A
ζ
D
B
88
88
g
88
Ta 8
8 / TA
mA
Ta
We have a truncated simplicial object:
T 2a
/
T A mT A / T 2 A
/
3
T mA
Ta
o T iA
mA
/
/
A
We now form a 2-categorical colimit, called the codescent object, of this truncated
simplicial object, and the result is the desired A0 . Alternatively, we can break this
up into bite-sized pieces. We first construct the coinserter of mA and T a: this is
the universal p : T A → A1 equipped with a 2-cell ρ : p.mA → p.T a. To give a
map A1 → B in T -Algs is equivalently to give a map f : A → B in K and a
2-cell f : b.T f → f a, without any coherence conditions. To capture the coherence
conditions, we have to perform a special sort of quotient, called a coequifier, which
universally makes equal a parallel pair of 2-cells. We’ll talk about 2-categorical
limits and colimits later.
If we used the “pseudo” version of weak morphisms, then we’d use a co-isoinserter
instead of an coinserter, which is the obvious analogue in which ρ is invertible.
A 2-CATEGORIES COMPANION
21
4.3. Consequences of the pseudomorphism classifier. Recall that we have
0
> A @@
~
>
@@q
~>
@@
~> ~>
@
/A
A
1
p
with qp = 1. It’s also true that pq ∼
= 1, so that this is an equivalence, and thus
A ' A0 in T -Alg, although generally not in T -Algs . If however q has a section s in
T -Algs , so that qs = 1, then s ∼
= 1, and q is an equivalence in T -Algs .
= p, so sq ∼
When q does have such a section, the algebra A is said to be flexible.
You can think of A0 as being a cofibrant replacement for A. We’ll see in Section 7.3 that there is a model structure on T -Algs for which A0 is a cofibrant
replacement of A. The weak equivalences are the strict morphisms which become
equivalences in T -Alg, or equivalently in K ; the cofibrant objects are precisely the
flexible algebras.
Exercise 4.3. If A is flexible, then any pseudo A
A → B.
B is isomorphic to a strict
The equivalence A ' A0 is a kind of coherence result for morphisms. There are
also coherence results for algebras. Consider the composite
T -Algs → T -Alg → Ps-T -Alg
To give a left adjoint is to construct a pseudo morphism classifier A0 ∈ T -Algs
not just for each strict T -algebra, but also for pseudo-T -algebras. This can still be
done; rather than a truncated simplicial object one has a truncated pseudosimplicial
object (some of the simplicial identities are satisfied only up to isomorphism), but
we can still form the codescent object A0 and obtain an isomorphism of categories
T -Alg (A0 , B) ∼
= Ps-T -Alg(A, B)
s
for any strict algebra B, natural in B with respect to strict maps. This time we
have a counit q : B 0 → B only when B is strict, and a unit A
A0 for any pseudo
algebra A. For a general pseudo algebra A, there seems no way to construct a
map from A0 back to A, and so no way to show that p is an equivalence. In some
cases, however p is an equivalence. In particular it is so if T preserves the relevant
codescent objects, since then one can construct the codescent object in K , and get
the inverse-equivalence down there. There are various other sufficient conditions
for this to work.
The existence of A0 for each pseudoalgebra A, along with the fact that the unit
A
A0 is an equivalence is sometimes called the “full coherence result”.
There are 2-monads for which not every pseudoalgebra is equivalent to a strict
one, but the only examples I know involve horrible 2-categories K . I don’t know
of an example satisfying the assumptions made in this section (K cocomplete and
T preserving α-filtered colimits).
5. Presentations for 2-monads
Presentations involve free gadgets and colimits. Both are defined in terms of a
universal property involving maps out of the constructed gadget. Why are these important in the case of 2-monads (or monads)? It turns out that one can use colimits
to build up 2-monads out of free ones exactly as one builds up algebraic structure
22
STEPHEN LACK
using basic operations, derived operations, and equations. Both the colimits and
the freeness will involve the world of strict morphisms of monads. Exactly what
this world might be is discussed below, but to start with we indicate why (strict)
maps out of a given monad are important.
5.1. Endomorphism monads. Let T be a monad on a complete category K .
Everything works without change for 2-categories, or indeed for V -categories. For
objects A, B ∈ K , the right Kan extension
K
~> CCC hA,Bi
~
CC
~ CC
~~
~
C!
~~
/K
1
A
B
can be computed as
hA, BiC = K (C, A) t B
where t means the cotensor, defined by
K (D, X t B) ∼
= Cat(X, K (D, B))
for a set (or category or object of V , as the case may be) X, and objects B and D
of K . The universal property of the right Kan extension implies in particular that
we have bijections of natural transformations.
T −→ hA, Bi
T A −→ B
This is starting to look like something you might want to do if T is a monad.
We have a natural “composition” and “identity” maps
hB, CihA, Bi −→ hA, Ci
1 → hA, Ai
which provide K with an enrichment over [K , K ] with internal-hom hA, Bi.
(Writing down where the composition and identity come from is a good exercise.)
Thus hA, Ai becomes a monoid in [K , K ]; that is, a monad. This can be regarded
as the monoid of endomorphisms of A in the [K , K ]-category K .
The important thing about this monad is that the bijection
T −→ hA, Ai
T A −→ B
restricts to a bijection
monad
T −→ hA, Ai
alg. str.
T A −→ A
between monad maps into hA, Ai and algebra structures on A.
This tells us that colimits of monads are interesting. For example, algebras for
S + T (coproduct as monads) are objects with an algebra structure for S and an
algebra structure for T , with no particular relationship between the two.
This is exactly like the endomorphism operad of an object, except that instead of
an object of n-ary operations for each n ∈ N, we have an object “C-ary operations”
hA, AiC = K (C, A) t A
for each object C ∈ K .
A 2-CATEGORIES COMPANION
23
We can play the same game with morphisms. First observe that hA, Bi is functorial (covariant in B, contravariant in A), and so for any f : A → B we can form
the solid part of
/ hB, Bi
β
T
hf,Bi
α
hA, Ai
/ hA, Bi
hA,f i
and now if we have monad maps α : T → hA, Ai and β : T → hB, Bi, then the
square commutes if and only if f is a strict map between the corresponding algebras.
It is at this point that we want to make things 2-categorical, and allow for pseudo
or lax morphisms. So suppose that K is a (complete) 2-category, and that T is a
2-monad on K . To give a 2-cell
/ TB
Tf
TA
a
A
f
f
b
/B
is equivalently to give a 2-cell
T
α
hA, Ai
/ hB, Bi
β
f˜
hA,f i
hf,Bi
/ hA, Bi.
The comma object
{f, f }`
c
hA, Ai
d
/ hB, Bi
hf,Bi
hA,f i
/ hA, Bi
is the universal diagram of this shape, so to give f˜ as above is equivalent to giving
a 1-cell ϕ : T → {f, f }` with dϕ = β and cϕ = α.
Now {f, f }` becomes a monad: this can be seen via a routine argument using
pasting diagrams; or one can get more sophisticated, and show that Lax(2, K ) is
enriched over [K , K ], and now regard {f, f }` as the endomorphism monoid. The
important thing is that
{f, f }`
d
/ hB, Bi
c
hA, Ai
are monad maps (although hA, Bi is not a monad), and that a map T → {f, f }` is
a monad map if and only if the corresponding (f, f ) is a lax T -morphism.
24
STEPHEN LACK
Of course there is also a pseudo version of this: use the isocomma object {f, f }
rather than the comma object {f, f }` ; this is the evident analogue in which the
2-cell is required to be invertible.
Thus we can work out the algebras and the (strict, pseudo, or lax) morphisms
for a monad just by looking at monad morphisms out of T , and that shows why
free monads and colimits of monads should be important.
Exercise 5.1. Describe the T -transformations in this way.
5.2. Pseudomorphisms of monads. In addition to strict monad maps, where
the good colimits live, there are also pseudo maps of monads. A pseudomorphism
of 2-monads on K is a 2-natural transformation, which preserve the multiplication
and unit in exactly the sense that strong monoidal functors preserve the tensor
product and unit of monoidal categories. Thus there are isomorphisms
i /
1>
T o
>> ∼
=
>>
> f
j >>
T o
m
∼
=
n
T2
S
f2
satisfying the same usual coherence conditions.
We have seen that to give an arbitrary map α : T → hA, Ai is equivalent to
giving a : T A → A in K , and that α is a strict map of monads if and only if a
makes A into a strict algebra; it turns out that to make α into a pseudomorphism
of monads
α:T
hA, Ai
is precisely equivalent to making
a : TA → A
into a pseudoalgebra.
5.3. Locally finitely presentable 2-categories. For a large 2-category K , the
2-category Mnd(K ) of 2-monads on K has all sorts or problems: its hom-categories
are large, it is not cocomplete, and free monads don’t exist. We shall therefore pass
to a smaller 2-category of 2-monads.
Assume that K is a locally finitely presentable 2-category. If you know what
a locally finitely presentable category is then this is just the obvious 2-categorical
analogue. If not, then here are some ways you could think about them:
• The formal definition (which you don’t need to know because I’m not going
to prove anything): a cocomplete 2-category with a small full subcategory
which is a strong generator and consists of finitely presentable objects.
• A 2-category which is complete and cocomplete and in which transfinite
arguments are more inclined to work then usual.
• A 2-category of all finite-limit-preserving 2-functors from C to Cat, where
C is a small 2-category with finite limits; you can take C to be Kfop where
Kf is the full subcategory of finitely presentable objects.
• Full reflective sub-2-categories of presheaf 2-categories which are closed
under filtered colimits.
• A 2-category which is complete and cocomplete, and is the free cocompletion under filtered colimits of some small 2-category (an Ind-completion).
A 2-CATEGORIES COMPANION
25
In fact you don’t need to suppose both completeness and cocompleteness:
for an Ind-completion, either implies the other.
Examples include the presheaf 2-category [A , Cat] for any small 2-category A ,
or CatX for any set X. The 2-category of groupoids is another example.
Once again, this is really an enriched categorical notion: there is a notion of
locally finitely presentable V -category, provided that V itself has a good notion
of finitely presentable object: more precisely, provided that V is a locally finitely
presentable category and the full subcategory of finitely presentable objects is closed
under the monoidal structure.
Because K is the free completion of Kf under filtered colimits, to give an
arbitrary 2-functor Kf → K is equivalent to giving a finitary (that is, filteredcolimit-preserving) 2-functor K → K . We write Endf (K ) for the monoidal
2-category of finitary endo(-2-)functors on K . Unlike [K , K ] this is locally small,
since Kf is small.
A 2-monad is said to be finitary if its endo-2-functor part is so. Then the 2category Mndf (K ) of finitary 2-monads on K is the 2-category of monoids in
Endf (K ), and the forgetful 2-functor U : Mndf (K ) → Endf (K ) does indeed
have a left adjoint, so in this world we do have free monads.
Moreover, the adjunction is monadic; there is a 2-monad on Endf (K ) for which
Mndf (K ) is the strict algebras and strict morphisms. We can drop down even
further to get
Mndf (K )
= G
H
F
a W
Endf (K )
G
G
a
U
V
}
[obKf , K ]
and go back up (along G) by left Kan extension along the inclusion obKf → K ).
The lower adjunction is also monadic, as indeed is the composite, although this
does not follow from monadicity of the two other adjunctions.
Thus Mndf (K ) is monadic both over Endf (K ) and over [obKf , K ], and the
choice of which base 2-category to work over affects what the pseudomorphisms
and pseudoalgebras will be. Dropping down one level, the transformations are 2natural ones, as in Section 5.2; while if we drop down the whole way, they will be
only pseudonatural.
The induced monads on [obKf , K ] are finitary, and so it follows that Endf (K )
and Mndf (K ) are themselves locally finitely presentable, and in particular are
complete and cocomplete. In fact slightly more is true, since the inclusion of
Mndf (K ) in Mnd(K ) has a right adjoint, and so preserves colimits. Mnd(K )
does not have colimits in general, but it does have colimits of finitary monads, and
these are finitary. Free monads on arbitrary endo-2-functors may not exist, but
free monads on finitary endo-2-functors do, and they are themselves finitary. This
is useful since the hA, Ai are not finitary, although we can use the coreflection of
Mnd(K ) into Mndf (K ) to obtain a finitary analogue.
Everything in this section remains true if you replace “finite” by some regular
cardinal α.
26
STEPHEN LACK
5.4. Presentations. The most primitive generator for a 2-monad is an object of
[obKf , K ]: a family (Xc )c∈obKf of objects of K , indexed by the objects of Kf .
This then generates a free 2-monad F X. What is an F X-algebra? A monad map
F X → hA, Ai
which is the same as
X → U hA, Ai.
This just means that for each c, we have
Xc → hA, Aic
which unravels to a functor
K (c, A) → K (Xc, A)
between hom-categories. Since K is cocomplete, this is the same as a map
X
K (c, A) · Xc −→ A
c
where K (c, A) · Xc denotes the tensor of Xc ∈ K by the category K (c, A). Thus
we can think of Xc as the “object of all c-ary operations”.
Example 5.2. Let K = Cat, so Kf is
assigns to every such c a category Xc of
(
1
Xc =
0
the finitely presentable categories, and X
c-ary operations. We take
c = 0, 2
otherwise
where 2 is the discrete category 1 + 1. Thus we have one binary operation and one
nullary operation. An F X-algebra is then a category A with maps as above. If Xc
is empty, then K (Xc, A) is terminal, so there’s nothing to do. In the other cases,
we get maps
A2 → A
when c = 2 and
A0 = 1 → A
when c = 0. This is the first step along the path of building up the 2-monad
for monoidal categories. The pseudo (or lax) morphisms can be determined using
{f, f } or {f, f }` : they will preserve ⊗ and I in the pseudo or the lax sense, as the
case may be, but without coherence conditions.
Example 5.3. Again let K = Cat, and let
(
2 c=1
Xc =
0 otherwise
Then an F X algebra is a category with a map
A → A2
in other words, a pair of maps with a natural transformation
A
&
8A.
This is an example in which Xc is not discrete. We say that X specifies a “basic
operation of arity 1 (unary) and type arrow”.
A 2-CATEGORIES COMPANION
27
In the case of monoidal categories, there are operations of “type arrow” providing
the associativity and unit isomorphisms, but I’ll take a different approach.
5.5. Monoidal categories. Actually, let’s forget about the units, just worry about
the binary operation. Then Xc is 1 if c = 2 and 0 otherwise, so an F X-algebra
is a category with a single binary operation. Then we have a (non-commutative)
diagram of 2-categories and 2-functors
F X-Algs
JJ
t
JJ Us
JJ
=
6
t
JJ
t
t
J%
ytt
/ Cat
Cat
Cat(3,−)
Us ttt
which act on an F X-algebra (C, ⊗) by
(C,
9 ⊗)
EEE
y
y
y
EE
y
EE
yy
y
E"
|y
/ C3
C
C
and now we have the two maps
⊗(⊗1)
C3
⊗(1⊗)
/
/C
which are natural in (C, ⊗), and so induce two natural transformations Cat(3, −)Us →
Us in the previous triangle. We can take their mates under the adjunction Fs a Us
to get 2-cells in
F X-Algs
JJ
t9
JJ Us
Fs ttt
JJ
⇑⇑
JJ
tt
t
J%
tt
/ Cat
Cat
Cat(3,−)
with two 2-cells in the middle. Note that Us Fs = F X is the monad, so that we
have two natural transformations
Cat(3, −) ⇒ F X,
which are morphisms of endofunctors. We can now construct the free 2-monad
HCat(3, −) on Cat(3, −) and the induced monad morphisms
κ1 , κ2 : HCat(3, −) ⇒ F X.
Consider now an F X-algebra (C, ⊗), and the corresponding monad map γ :
F X → hC, Ci. Then (C, ⊗) is strictly associative if and only if γκ1 = γκ2 , while
to give an isomorphism ⊗(1⊗) ∼
= ⊗(⊗1) is equivalent to giving an isomorphism
γκ1 ∼
= γκ2 . In the 2-category Mndf (K ) construct the universal map ρ : F X → S
equipped with an isomorphism ρκ1 ∼
= ρκ2 : this is called a co-iso-inserter, and it’s
a (completely strict) 2-categorical colimit, which we’ll meet later on.
Now, an S-algebra is a category C with a functor ⊗ : C 2 → C and a natural
isomorphism α : ⊗(1⊗) ∼
= ⊗(⊗1). You can also write down what it means to be a
pseudo or lax morphism of such algebras, and it’s what you want it to be; the tensorpreserving isomorphisms must be compatible with the associativity constraints.
28
STEPHEN LACK
To do the coherence condition, we have a pair of 2-cells
ρκ1
HCat(4, −)
(
α1 ⇓ ⇓α2
6S
ρκ2
which encode the two isomorphisms that one wants to make equal. Now we form the
coequifier q : S → T , in the category of monads, of these two 2-cells: the universal
map with the property that qα1 = qα2 of making them equal.
Then the 2-category T -Alg is the 2-category of “semigroupoidal categories” and
strong morphisms (we can get the strict and lax morphisms in the obvious way
too). All this follows from the universal property of the monad T .
Often, as here, we build up structure in a particular order, starting with the
operations of type object, then those of type arrow or isomorphism, and finally
impose equations on these arrows or isomorphisms.
5.6. Terminal objects. Consider the structure of category with terminal object.
This is a baby example, but you can do any limits you like once you understand
this example.
How do you say algebraically that a category A has a terminal object? You give
an object
t
1 −→ A
with a natural transformation
1
/A
A?
?
??

?? τ 
t
! ??

1
such that the component
t
1
/A
??
??
??
?
/A
?

τ 

1
1
of τ at t is the identity. This last condition plus naturality of τ guarantees that
τ a : a → t is the only map from a to t, and so that t is terminal.
Let’s give a presentation for it. First we have the nullary operation t, which
takes the form
Cat(0, A) → Cat(1, A)
or equivalently
Cat(0, A) · X0 → A
where X0 = 1, or equivalently
X
Cat(c, A) · Xc → A
c
where now Xc is 0 unless c = 0. Thus an object A with nullary operation t : 1 → A
is precisely an F X-algebra, where
(
1 if c = 0
Xc =
0 otherwise
A 2-CATEGORIES COMPANION
29
For any F X-algebra (A, t), there are two canonical maps from A → A, given by
A II
II
I
! I$
1
1
/: A
uu
u
u
uu t
and these are clearly natural in (A, t); in other words, they define a pair of natural
transformations from Us : F X-Algs → Cat to itself. Taking mates under the
adjunction Fs a Us gives a pair of natural transformations 1Cat → Us Fs . Now
Us Fs is just F X, so forming the free monad H1 on the identity 1Cat , we get a pair
of monad maps κ1 , κ2 : H1 → F X. We now form the coinserter ρ : F X → S of κ1
and κ2 . This is another 2-categorical colimit; it is the universal ρ equipped with a
2-cell ρκ1 → ρκ2 . An S-algebra is now an A equipped with an object t : 1 → A,
and a natural transformation τ : 1A → t◦!, as in our earlier description of terminal
objects. Finally one can construct a suitable coequifier q : S → T to obtain the
2-monad T for categories with terminal objects.
Here’s a different presentation: it starts as before by putting in a nullary operation
t
Cat(0, A) −→ Cat(1, A)
but then adds a unary operation of type arrow:
Cat(1, A) −→ Cat(2, A)
τ
which specifies two endomorphisms of A and a natural transformation between
them:
f
A
τ
&
8 A.
g
Later we’ll introduce equations f = 1, g
To specify t and τ , define


1
Xc = 2


0
= t◦!, and τ t = id.
if c = 0
if c = 1
otherwise
so that F X-algebra structure an a category A amounts to
X
Cat(c, A) · Xc → A
c
or equivalently t : 1 → A and τ : f → g : A → A.
Now we turn to the equations. Consider the (non-commuting) diagram
F X-Algs
JJ
t
JJ Us
JJ
=
6
t
JJ
t
t
J%
t
yt
/ Cat
Cat
2·−+2
Us ttt
30
STEPHEN LACK
which acts on an F X-algebra (A, t, τ ) by
(A,
t, τ )
III
u5
u
II
uu
II
u
II
uu
u
I$
zuu
/
A+A+2
A
A.
There is a map α(A,t,τ ) : A + A + 2 → A whose components are f : A → A,
g : A → A, and the functor 2 → A corresponding to τ ◦ t. This is natural in
(A, t, τ ).
There is another map β(A,t,τ ) : A + A + 2 → A whose components are 1 : A →
A, t◦! : A → A, and the functor 2 → A corresponding to the identity natural
transformation on t. Once again this is natural in (A, t, τ ).
A category with terminal object is precisely an F X-algebra (A, t, τ ) for which
α(A,t,τ ) = β(A,t,τ ) .
Now α and β live in the diagram
F X-Algs
JJ
t
JJ Us
JJ
α⇑ ⇑β
JJ
tt
t
J%
ytt
/ Cat
Cat
E
Us ttt
where EC = C + C + 2, and we can take their mates under the adjunction Fs a Us
to obtain natural transformations
α0 , β 0 : E → Us Fs
and now Us Fs is the monad F X, so there are induced monad maps
α, β : HE → F X
from the free monad HE on E, and the required 2-monad T for categories-withterminal object is obtained as the coequalizer
α
HE
/
/ FX
q
/ T.
β
Remark 5.4. Whichever approach we take, the algebras will be the categories with
a chosen terminal object. This may seem strange, but is not really a problem. The
strict morphisms preserve the chosen terminal object strictly, which is probably not
what we really want, but the pseudo morphisms preserve it in the usual sense.
5.7. Bicategories. There are two reasons for including this example: first of all
it’s a fairly easy case with K 6= Cat, and second it’s important for 2-nerves. I
won’t give all the details.
Let K = Cat-Grph, the 2-category of category-enriched graphs. A Cat-graph
consists of a set of things G, H, . . . and hom-categories G (G, H) ∈ Cat. (Of course
one could do this for any V in place of Cat.) A morphism is a function G 7→ F G
of objects, along with functors G (G, H) → H (F G, F H) between hom-categories.
One might hope that the 2-cells would be some sort of natural transformations, but
since Cat-graphs have no composition law, there is no way to assert that a square
A 2-CATEGORIES COMPANION
31
in a Cat-graph commutes, and so no way to state naturality. Instead, we use a
special sort of lax naturality. We only allow 2-cells
F
G
(
7H
F0
to exist when F = F 0 on objects, and then the 2-cell consists of natural transformations
F
G (G, H)
!
= H (F G, F H)
F0
on all hom-categories.
Now, given a Cat-graph, what do you need to do to turn it into a bicategory?
To start with, you have to give compositions
G (H, K) × G (G, H) −→ G (G, K).
Let comp and arr be the Cat-graphs · → · → · and · → · (no 2-cells). Then
X
K (comp, G ) =
G (H, K) × G (G, H).
G,H,K
K (arr, G ) =
X
G (G, K)
G,K
so if we define
(
arr
Xc =
0
if c = comp
otherwise
then an F X-algebra structure on G amounts to a map
X
K (c, G ) · Xc → G
c
and so to a map
M:
X
G (H, K) × G (G, H) →
G,H,K
X
G (G, K).
G,K
We need to make sure that the restriction
MG,H,K : G (H, K) × G (G, H) →
X
G (G, K)
G,K
to the (G, H, K)-component lands in the (G, K)-component: this can be done by
constructing a quotient of F X.
Define
(
ob if c = comp
Yc=
0
otherwise
where ob denotes the Cat-graph · (no 1-cells or 2-cells). An F Y -algebra structure
on a Cat-graph G is a map
X
X
G (H, K) × G (G, H) →
1
G,H,K
where 1 denotes the terminal category.
G
32
STEPHEN LACK
Suppose now that (G , M ) is an F X-algebra. There are many induced F Y algebra structures on G ; in particular, there are the following two:
P
P P
P
M
G!
/
/
G,H,K G (H, K) × G (G, H)
G
K G (G, K)
G1
P
P
G,H,K
G (H, K) × G (G, H)
injG
/1
!
/
P
G
1
Each is functorial, and so each induces a monad map F Y → F X; we form their
coequalizer q1 : F X → S1 , and now an S1 -algebra
Pis a Cat-graph G equipped with
a composition M such that MG,H,K lands in
K G (G, K). A further quotient
forces MG,H,K to land in G (G, K) as desired.
One now introduces an associativity isomorphism. This has the form of a map
K (triple, G ) → K (iso, G )
where triple is the Cat-graph · → · → · → · and iso is ·
∼
=
#
; · . There are also left
and right identity isomorphisms, and various coherence conditions to be encoded,
but I’ll leave all that as an exercise. The result of the exercise is:
• An algebra is a bicategory.
• A lax morphism is a lax functor.
• A pseudo morphism is a pseudo functor.
• A strict morphism is a strict functor.
• A 2-cell is an icon. This is an oplax natural transformation (which we
haven’t officially met yet) for which the 1-cell components are identities.
ICON stands for “Identity Component Oplax Natural-transformation”. An
icon F → G can exist only if F and G agree on objects, in which case it
consists of a 2-cell
FA
GA
Ff
__ +3
Gf
FB
GB
for each f : A → B in G , subject to conditions expressing compatibility with
respect to composition of 1-cells and identities, and naturality in f with respect
to 2-cells. In the case of one-object bicategories these are precisely the monoidal
natural transformations.
These icons are just nice enough to give us a 2-category of bicategories. In
general, lax natural transformations between lax functors can’t even be whiskered
by lax functors — the composite
/
>
/
isn’t well-defined. In the pseudo case it is defined, but not associative, and so we
are lead into the world of tricategories. But with just icons, we do get a 2-category,
which is moreover the category of algebras for the 2-monad just described.
For example, in this 2-category, it’s true that every bicategory is equivalent (in
the 2-category) to a 2-category; this works because in replacing a bicategory by
a biequivalent 2-category you don’t have to change the objects of the bicategory.
A 2-CATEGORIES COMPANION
33
The 2-category NHom of bicategories, normal homomorphisms, and icons, is a
full sub-2-category of the the 2-category [∆op , Cat] of simplicial objects in Cat,
via a “2-nerve” construction. In order to deal with normal homomorphisms (which
preserve identities strictly) rather than general ones, it’s convenient to start with
reflexive Cat-graphs rather than Cat-graphs.
The choice of direction of the 2-cell in lax transformations and oplax transformations goes back to Bénabou. It seems that the oplax transformations are generally
more important than the lax ones.
5.8. Cartesian closed categories. The comments in this section apply equally
to monoidal closed categories, symmetric monoidal closed categories, and toposes.
There is no problem constructing a monad for categories with finite products,
similarly to the constructions given above. When we come to the closed structure,
however, things are not so straightforward. The internal hom is a functor
Aop × A → A
and we’re not allowed to talk about Aop the way we’re doing things: our operations
are supposed to be of the form Ac → A. How can we deal with this?
In fact, it’s a theorem that cartesian closed categories don’t have the form T -Alg
for a 2-monad T on Cat. What you can do, however, is change the base 2-category
K to the 2-category Catg of categories, functors, and natural isomorphisms. Recall
that Cat(2, A) is just A × A, but in Catg (2, A) we have only Aiso × Aiso . The
internal-hom does give us a functor
Aiso × Aiso
(a, b)
/ Aiso
/ [a, b]
which has the form
Catg (2, A) −→ Catg (1, A)
since we can turn around an isomorphism in the first variable to make everything
covariant. This gives a new problem; the product is now only given as a functor
Aiso × Aiso → Aiso , we have to put in the rest of the functoriality separately “by
hand”, using an operation
/ Catg (2, A)
Catg (2 + 2, A)
(f : a → a0 , g : b → b0 )
/ (f × g : a × b → a0 × b0 )
subject to various equations. You also have to relate the product to the internal
hom.
Any 2-monad on Cat induces monads on Catg and on the 1-category Cat0 (since
things are stable under change of base enriching category, categories to groupoids
to sets). But at each stage, to present the same structure becomes harder. In the
groupoid enriched stage we can still talk about pseudomorphisms, although at this
stage every lax morphism is pseudo; by the time we get to the Set-enriched stage
there is no longer any genuine pseudo notion at all — everything is strict.
34
STEPHEN LACK
5.9. Diagram 2-categories. The first version of this is not really an example
of a presentation at all, since the 2-monad pops out for free. Let C be a small
2-category, and consider the 2-category [C , Cat] of (strict) 2-functors, 2-natural
transformations, and modifications. This is the Cat-enriched functor category.
The forgetful 2-functor has both adjoints
[C , Cat]
^
@
a
a
[obC , Cat]
given by left and right Kan extension. The existence of the right adjoint tells us
that the forgetful functor preserves all colimits. In this case Us is strictly monadic
as is easy to prove with the enriched Beck’s theorem. The induced monad T then
preserves all colimits, and we can write, using the Kan extension formula,
X
(T X)c =
C (d, c) · Xd.
d
It’s now a long, but essentially routine, exercise to check that
•
•
•
•
pseudo T -algebras are pseudo-functors,
lax algebras are lax functors,
pseudo morphisms are pseudo-natural transformations,
etc.
When you write down the coherence conditions for a lax morphism it will tell you
more than is in the definition of a lax functor: it will also include a whole lot of
consequences of the definition.
Now let C be a bicategory. If we tried the same game, we wouldn’t get a 2monad, since the associativity of the multiplication for the monad corresponds to
the associativity of composition in C , so we’d just get a pseudo-monad. We could
just go ahead and do this, but we’ve been avoiding pseudo-monads, and there is an
alternative. One can give a presentation for a 2-monad T on [obC , Cat] whose
• (strict) algebras are pseudofunctors C → Cat,
• pseudomorphisms of algebras are pseudonatural transformations,
• etc.
You start with a family (Xc )c∈obC , then introduce operations
C (c, d) × Xc → Xd
and so on. The target doesn’t really need to be Cat, although it would need to be
cocomplete.
6. Limits
We’ll begin with some concrete examples, looking in particular at limits in T -Alg,
for a finitary 2-monad T on a locally finitely presentable 2-category K (you could
get by with much less for most of this).
A 2-CATEGORIES COMPANION
35
6.1. Terminal objects. Let’s start with something really easy: terminal objects.
Let 1 be terminal in K ; we have a unique map T 1 → 1, making 1 a T -algebra, and
then for any T -algebra (A, a) we have a unique ! : A → 1, and
/ T1
T!
TA
A
/1
!
commutes strictly, so there’s a unique strict algebra morphism A → 1. Moreover,
by the 2-universal property of 1, there’s a unique isomorphism in the above square,
which happens to be an identity; thus there is only one pseudo morphism as well
(which happens to be strict). A similar argument works for endomorphisms of this
morphism; thus
T -Alg((A, a), (1, !)) ∼
=1
so (1, !) is a terminal object in T -Alg.
6.2. Products. Similarly for products: given a product A × B in K , there is an
obvious map ha, bi as in
T (A × B) → T A × T B → A × B
which makes A × B into a T -algebra (exactly as for ordinary monads: nothing
2-categorical going on here). The point is that if we have pseudo morphisms
/ TB
/ TA
TC
TC
c
C
∼
=
a
c
/A
C
∼
=
b
/B
we get a unique induced pseudo morphism
/ T (A × B)
TC
c
C
∼
=
ha,bi
/ A×B
and indeed there is a natural isomorphism (of categories)
T -Alg(C, A × B) ∼
= T -Alg(C, A) × T -Alg(C, B).
Thus A × B is a product in T -Alg in the strict Cat-enriched sense.
Note that the projections A × B → A and A × B → B are actually strict maps,
by construction. Moreover, they jointly “detect strictness”: a map into A × B is
strict if and only if its composites into A and B are strict. This is a useful technical
property.
Actually, we didn’t really need to check anything, since we’ve already seen that
T -Algs ,→ T -Alg has a left adjoint, hence preserves all limits, and in the case
of terminal objects and products the diagram of which we are taking the limit
consists only of objects, so already exists in the strict world. (On the other hand,
the explicit argument works for any 2-monad on any 2-category with the relevant
products, whereas the adjunction needs a transfinite argument, and much stronger
assumptions on T and K .)
36
STEPHEN LACK
6.3. Equalizers. Now let’s look at equalizers. Here it’s different, because the
morphisms whose equalizer we seek may not be strict. If they are, then the equalizer
exists in T -Algs and is preserved, but if they aren’t, the adjunction doesn’t help.
In fact, in general equalizers of pseudo morphisms need not exist.
For example, let T be the 2-monad on Cat for categories with a terminal object.
Let 1 be the terminal category and let I be the free-living isomorphism, consisting
of two objects and a single isomorphism between them. Clearly both categories
have a terminal object, and both inclusions are pseudo morphisms. But any functor
which equalizes them has to have empty domain, and no category with an empty
domain has a terminal object.
6.4. Equifiers. Thus T -Alg is not complete, but we can look at some of the limits
that it does have. Consider a parallel pair of 1-cells in T -Alg with a parallel pair
of 2-cells between them:
f
A
α⇓ ⇓β
)
5 B.
g
The equifier of these 2-cells, is the universal 1-cell k : C → A with αk = βk.
Here universality means that K (D, C) is isomorphic (not just equivalent) to the
h
category of morphisms D −→ A with αh = βh. Equifiers do lift from K to T -Alg:
if (A, a) and (B, b) are T -algebras, (f, f ) and (g, g) are T -morphisms, and α and β
are T -transformations, then the composites
Tf
TC
Tk
Tf
)
T α 5 T B
/ TA
b
/B
=
TC
Tk
)
T β 5 T B
/ TA
Tg
b
/B
Tg
are equal. Paste the isomorphism g : b.T g ∼
= ga on the bottom of each side and the
isomorphism f : f a ∼
= b.T f on the top, and use the T -transformation condition for
α and β to get the equation
f
TC
Tk
/ TA
a
/A
g
α
(
6B
f
=
TC
Tk
/ TA
a
/A
β
(
6B
g
and now by the universal property of the equifier C there is a unique c : T C → C
satisfying kc = a.T k. Two applications of the universal property show that c makes
C into a T -algebra, and so clearly k becomes a strict T -morphism (C, c) → (A, a).
Further judicious use of the universal property shows that k : (C, c) → (A, a) is
indeed the equifier in T -Alg.
Observe that once again, the projection map k of the limit is actually a strict
map, and detects strictness of incoming maps.
Why does the analogous argument for equalizers fail? Given pseudo morphisms
(f, f ) and (g, g) from (A, a) to (B, b), we could form the equalizer k : C → A of
f and g, and then hope to make C into a T -algebra using the universal property
of C, but we’d need to show that f a.T k = ga.T k. All we actually know is that
c.T f.T k = c.T g.T k while c.T f.T k ∼
= f a.T k and c.T g.T k ∼
= ga.T k, which just isn’t
good enough.
The moral is that in forming limits in T -Alg, we can ask for existence or invertibility of 2-cells, and equations between them, but we can’t generally force equations
between 1-cells.
A 2-CATEGORIES COMPANION
37
6.5. Inserters. There is a sort of lax version of an equalizer, called an inserter.
Rather than making 1-cells equal, you put a 2-cell in between them. The inserter
of a parallel pair of arrows f, g : A → B is the universal k : C → A equipped with
a 2-cell κ : f k → gk. More precisely, the universal property states that K (D, C)
should be isomorphic to the category whose objects are morphisms ` : D → A
equipped with a 2-cell λ : f ` → g`, and whose morphisms (`, λ) → (m, µ) are
2-cells
`
D
α
&
8A
m
such that
f`
λ
/ g`
gα
fα
fm
κ
/ gm
commutes.
Once again, inserters in K lift to T -Alg, where they have strict projections and
detect strictness. Given a pair
(f,f )
(A, a)
(
6
(B, b)
(g,g)
of pseudo morphisms, we construct the inserter (k : C → A, κ : f k → gk) of f and
g in K , and want to make it an algebra. We need a 2-cell f.a.T k → g.a.T k to
induce c : T C → C, so we follow our nose:
b.T κ
f.a.T k ∼
= b.T f.T k −→ b.T g.T k ∼
= g.a.T k
This thing then must be κc for a unique c, by the universal property of the inserter
in K . Now check that c makes C into an algebra, and so on; everything goes
through just as before.
Observe that an inserter in a (2-)category with no non-identity 2-cells is just an
equalizer.
6.6. PIE-limits. Thus T -Alg has Products, Inserters, and Equifiers, and many
important types of limit can be constructed out of these. A limit which can be so
constructed is called a PIE-limit, so clearly T -Alg has all PIE-limits, and equally
clearly equalizers are not PIE-limits. Some positive examples are:
• iso-inserters, which are inserters where we ask the 2-cell to be invertible.
Insert 2-cells in each direction, then equify their composites with identities.
(Of course you can’t go the other way: iso-inserters don’t suffice to construct
inserters.)
• inverters, where we start with a 2-cell α and make it invertible: we want
the universal k such that αk is invertible. Insert something going back the
other way, then equify composites with the identities.
• cotensors by categories. Cotensors by discrete categories can be constructed
using products. Any category can be constructed from discrete ones using
coinserters (to add morphisms) and coequifiers (to specify composites). So
38
STEPHEN LACK
cotensors by arbitrary categories can be constructed from cotensors by discrete categories using inserters and equifiers.
The dual (colimit) notions of some of these were important in giving presentations of monads. The dual of inverter is the coinverter. The coinverter of a 2-cell
α : f → g : A → B is the universal q : B → C with qα invertible. In Cat, this
is just the category of fractions B[Σ−1 ], where Σ consists of all arrows in B which
appear as components of α. Of course the dual of cotensor is tensor, not cocotensor!
6.7. Weighted Limits. In this section we briefly review the general notion of
weighted limit, before turning in the next section to the case V = Cat, where we
shall see how the various examples of the previous section arise.
Let S : C → K be a functor between, say, ordinary categories. The limit is
supposed to be defined by the fact that
K (A, lim S) ∼
= Cone(A, S)
where the right hand side is the set of cones under S with vertex A. This is typically
defined as the hom-set [C , K ](∆A, S), where ∆A denotes the constant functor at
A, but it can also be expressed as [C , Set](∆1, K (A, S)). It is this last description
of cones which forms the basis for the generalization to weighted limits; we’re going
to replace ∆1 by some more general functor C → Set.
Example 6.1. No one really uses this in practice, but it’s useful to think about,
and motivates the name “weighted” in “weighted limit”. Let C = 2 have two
objects, so a functor S : C → K is a pair of objects B and C, and a weight is a
functor J : C → Set, say it sends one to 2 and the other to 3. Then
[C , Set](J, K (A, S))
consists of functions 2 → K (A, B) and 3 → K (A, C), or equivalently two arrows
A → B and three arrows A → C, so that the “weighted product” is B 2 × C 3 .
For general V , we start with V -functors S : C → K and J : C → V and consider
[C , V ](J, K (A, S)).
If this is representable as a functor of A, the representing object is called the Jweighted limit of S and written {J, S}. Thus we have a natural isomorphism
K (A, {J, S}) ∼
= [C , V ](J, K (A, S)).
which defines the limit.
Exercise 6.2. If K = V , then {J, S} is the internal hom [C , V ](J, S).
When V = Set, weighted limits don’t give you any new limits: if K is an
ordinary category which is complete in the usual sense of having all conical limits
(J = ∆1), then it also has all weighted limits. More precisely, for any weight
J : C → Set and any diagram S : C → K , there is a category D and a diagram
R : D → K , such that the universal property of {J, S} is precisely the universal
property of the usual limit of R.
But the weighted ones are more expressive, so it’s still useful to think about them.
In particular, you might want to talk about all limits indexed by a particular weight
J : C → Set; this class is not so easy to express using only conical limits.
When V 6= Set it’s not longer true that all limits can be reduced to conical ones.
But if you have all conical limits and cotensors, you can construct all weighted
limits.
A 2-CATEGORIES COMPANION
39
Remark 6.3. There is a slight subtlety here. In the case V = Set, the conical limit
of a functor S : C → K is just the limit of S weighted by ∆1 : C → Set. But
for a V -category C , the “constant functor at 1” (from C to V ) is usually not what
you want to look at, and indeed may fail to exist. What you really want, to get the
right universal property, is the constant functor ∆I at the unit object I of V . But
even this may not exist, unless C is the free V -category on an ordinary category B.
So this is the right general context for conical limits in enriched category theory.
That’s all I want to say about general V .
6.8. Cat-weighted limits.
Example 6.4 (Inserters). Let C be the 2-category · ⇒ ·, so S is determined by a
parallel pair of arrows A ⇒ B. The weight J : S → Cat sends it to (1 ⇒ 2). Then
a natural transformation J → K (C, S) gives us 1 → K (C, A), hence an arrow
h : C → A, and 2 → K (C, B), hence
u
C
β
&
8B.
v
But by naturality we have u = f h and v = gh, so the data consists of 1-cell
h : C → A and a 2-cell β : f h → gh.
This is the 1-dimensional aspect of the universal property, which characterizes
the 1-cells into C; there is also a 2-dimensional aspect characterizing the 2-cells,
since the limit is defined in terms of an isomorphism of categories, not just a
bijection between sets. In general, this 2-dimensional aspect must be checked, but
if the 2-category K should admit tensors, the 2-dimensional aspect follows from
the 1-dimensional one. Similar comments apply to all the examples.
Example 6.5 (Equifiers). Here, our 2-category C is
(
α⇓ ⇓β
6
and our weight is
1
(
⇓⇓
62
in which α and β get mapped to the same 2-cell in Cat.
Example 6.6 (Comma objects). C is the same shape for pullbacks
/
and J is
1 .
1
0
1
/2
There is no 2-cell in C , since we don’t start with a 2-cell, we only add one universally.
40
STEPHEN LACK
Example 6.7 (Inverters). Recall, this is where we start with a 2-cell and universally
make it invertible. Then C is
'
7
and J is
1
'
7I
where I is the “free-living isomorphism” · ·.
6.9. Colimits. Colimits in K are limits in K op . That’s really all you have to say,
but I should show you the notation. As usual, we rewrite things so as to refer to
K . In this case, it’s also convenient to replace C by C op , so that we start with
S: C → K
J : C op → V
and now the weighted colimit is written J ? S and defined by a natural isomorphism
K (J ? S, A) ∼
= [C op , V ](J, K (S, A)).
One form of the Yoneda lemma says that
J∼
=J ?Y
where Y : C → [C op , V ] is the Yoneda embedding and J : C op → V is arbitrary.
Here’s an application. Suppose you have some “limit-notion” which you know in
advance is a weighted limit, but you don’t know what the weight is. Thus you know
{J, S} given S, but you don’t know J itself. Consider the version of the Yoneda
embedding Y : C → [C , V ]op and take its “limit”, for the notion of limit we’re
interested in; equivalently, take the relevant colimit of Y : C → [C op , V ]. This
is J ? Y for our as yet unknown J; but by the Yoneda lemma this J ? Y is itself
the desired weight. This can be used to calculate the weights for all the concrete
examples of Cat-weighted limits discussed here.
6.10. Pseudolimits. Now we are interested in
K (A, pslimS) ∼
= Ps(C , Cat)(∆1, K (A, S)).
where Ps(A , B) is the 2-category of 2-functors, pseudonaturals, and modifications
from A to B. The right side is what we mean by a pseudo-cone. Note that this is
still an isomorphism of categories, not an equivalence.
Example 6.8 (Pseudopullbacks). Again we take C to be
/
A pseudo-cone then consists of
/
∼
=
∼
=
/ with isomorphisms in each triangle. We have made the cones commute only up to
isomorphisms, but the universal property and factorizations are still strict. Note
A 2-CATEGORIES COMPANION
41
that the pseudopullback is equivalent (not isomorphic) to the isocomma object (assuming both exist). In the latter, we specify f a ∼
= gb without specifying the middle
diagonal arrow. Of course, we can take it to be f a, or gb, so we get ways of going
back and forth.
Pseudopullbacks are not in general equivalent to the pullback, although it is
possible to characterize when they are [14]. This situation is entirely analogous
to homotopy pullbacks, and indeed it can be regarded as a special case, via the
“categorical” Quillen model structure on Cat (see Section 7).
Again, given a weight J : C → Cat, the weighted pseudolimit is defined by
K (C, {J, S}ps ) ∼
= Ps(C , Cat)(J, K (C, S)).
I don’t really want to do any examples of this one, I want to do some general
nonsense instead.
Recall that Ps(C , Cat) = T -Alg for a 2-monad T on [obC , Cat], while T -Algs =
[C , Cat], so
J : [C , Cat] → Ps(C , Cat)
has a left adjoint Q, with QJ = J 0 . Thus
Ps(C , Cat)(J, K (C, S)) ∼
= [C , Cat](J 0 , K (C, S))
which just defines the universal property for the J 0 -weighted limit. In other words,
pseudolimits are not some more general thing, but a special case of ordinary (weighted)
limits. Thus we say that a weight “is” a pseudolimit if it is J 0 for some J.
Remark 6.9. This sort of phenomenon is common. Recall, for example, that pseudoalgebras for monads are strict algebras over a cofibrant replacement monad. Thus
talking about things of the form Ps-T -Alg is actually less general than things of
the form T -Alg, since everything of the former form has the latter form, but not
conversely.
6.11. PIE-limits. Recall that these are the limits constructible from products,
inserters, and equifiers. We can now make this more precise. A weight J : C → Cat
is a (weight for a) PIE-limit if and only if the following conditions hold:
• any 2-category K with products, inserters, and equalizers has J-weightedlimits;
• any 2-functor F : K → L which preserves products, inserters, and equalizers (and for which K has these limits) also preserves J-weighted limits.
There is a characterization of such weights. Given a 2-functor J : C → Cat,
first consider the underlying ordinary functor J0 : C0 → Cat0 obtained by throwing
away all 2-cells. Now compose this with the functor ob : Cat0 → Set which throws
away the arrows of a category, leaving just the set of objects. This gives a functor
j : C0 → Set. Then J is flexible if and only if j is a coproduct of representables;
and j will be a coproduct of representables if and only if each connected component
of the category of elements of j has an initial object.
Proposition 6.10. Pseudolimits are PIE-limits.
For a general T -algebra A, the pseudomorphism classifier A0 was constructed
from free algebras using coinserters and coequifiers. Thus for a general weight
J : C op → Cat we can construct J 0 from “free weights”, using coinserters and
coequifiers. Free weights, in this context, are coproducts of representables, thus J 0
42
STEPHEN LACK
can be constructed from representables using coproducts, coinserters, and coequifiers. Now a limit weighted by a representable C (C, −) is given by evaluation at
C; a limit weighted by a coproduct of representables is given by the product of the
evaluations; a limit weighted by a coinserter of coproducts of representables is given
by an inserter of products of evaluations, and so on. This is part of a general result,
not needed here, that colimits of weights give iterated limits, as in the formula
∼ {J, {H, S}}
{J ? H, S} =
reminiscent of a tensor-hom situation. In other words, the 2-functor
{−, S} : [C , Cat]op → K ,
sending a weight J to the J-weighted limit of S, sends colimits in [C , Cat] to limits
in K . In any case we can conclude in the current context that J 0 -weighted limits
are PIE-limits.
The converse is false: for example inserters are not pseudolimits. Neither are
iso-comma objects, although they’re pretty close (as we saw above).
Remember that T -Alg had all PIE-limits. It therefore has all pseudolimits as
well. But consider the class of all limits (weights) which are equivalent (in [C , Cat],
so that the equivalences are 2-natural) to pseudolimits. It is not the case that T -Alg
has all of those limits. So equivalence of limits is not always totally trivial.
For example, consider splitting of idempotent equivalences, which seems like a
very benign thing to do. If we split an idempotent equivalence
TA
a
T
Te
∼
=
e
/ TA
/T
a
in T -Alg, we won’t necessarily get a T -algebra back, only a pseudo-algebra.
As an example, let C be a non-strict monoidal category, and Cst its strictification.
Then there is an idempotent equivalence on Cst , which when split, gives C .
6.12. Bilimits. I’m going to write down all the same symbols, but they’ll just
mean different things! So now C and K are bicategories, while S : C → K and
J : C → Cat are now homomorphisms (pseudofunctors). The weighted bilimit is
defined by an equivalence
K (C, {J, S}b ) ' Hom(C , Cat)(J, K (C, S)).
Now our limits are determined only up to equivalence, instead of up to isomorphism.
In the case when C and K are 2-categories and J and S are 2-functors, then
the right hand side is equal to the right hand side for pseudolimits, just by definition (since Ps(C , Cat) ,→ Hom(C , Cat) is locally an isomorphism). Thus every
pseudolimit is a bilimit.
On the other hand, if just K is a 2-category, then you can replace C by a 2category C 0 such that homomorphisms out of C are the same as 2-functors out of
C 0 . Now for any A ∈ K we have
˜ K (S̃, A))
Hom(C op , Cat)(J, K (S, A)) ' Ps(C op , Cat)(J,
where J˜ and S̃ are the 2-functors corresponding to J and S. Thus a 2-category
with all pseudolimits also has all bilimits.
A 2-CATEGORIES COMPANION
43
As we shall see in the following section, the converse is false: there are 2categories with bilimits which do not have pseudolimits, so the definition of pseudolimit is logically harder to satisfy than that of bilimit. On the other hand in
concrete examples it is often much easier to verify the definition using pseudolimits
than bilimits. This is certainly the case for pseudolimits in T -Alg. It’s also the case
for the opposite of the 2-category of Grothendieck toposes.
6.13. Bilimits and bicolimits in T -Alg. Suppose once more that K is locally
finitely presentable and T is finitary, and consider the 2-category T -Alg. It has
PIE-limits, as we saw, and so has pseudo-limits, and so has bilimits. So from a
bicategorical perspective, we have all the limits we might want.
T -Alg also has bicolimits, although not in general pseudocolimits or PIE-colimits.
Thus T -Algop is an example of a 2-category with bilimits but not pseudolimits. Just
as in the case of ordinary monads, the (bi)colimits in T -Alg are not constructed as
in K .
The existence of bicolimits follows from:
Theorem 6.11. Suppose we have a 2-functor G : T -Alg → L such that the
composite GJ in
J
G
T -Algs −→ T -Alg −→ L
has a left adjoint F . Then JF is left biadjoint to G.
We start with a left 2-adjoint F to GJ but end up with only a left biadjoint to
G. Here’s the idea of the proof. The biadjunction amounts to a (pseudonatural)
equivalence
T -Alg(JF L, A) ' L (L, GA).
Since T -Algs and T -Alg have the same objects, we may write A as JA. Now the
adjunction F a GJ gives an isomorphism of categories
T -Alg (F L, A) ∼
= L (L, JGA)
s
so it suffices to show that
T -Algs (F L, A) ' T -Alg(JF L, JA)
which in turn amounts to the fact that every pseudomorphism from F L to A is
isomorphic to a strict one. This will hold if we know that (F L)0 ' F L. Writing Q
for the left adjoint to J : T -Algs → T -Alg, we have a pseudomorphism p : JF L
JQJF L (unit of Q a J), and a map n : L → GJF L (unit of F a GJ), so we can
form the composite
L
n
/ GJF L
Gp
/ GJQJF L
and the corresponding strict map
FL
r
/ QJF L
under the adjunction F a GJ provides the desired inverse-equivalence to q :
QJF L → F L (the counit of Q a J).
Corollary 6.12.
(i) T -Alg has bicolimits;
(ii) for any monad morphism f : S → T , the induced 2-functor f ∗ : T -Alg →
S-Alg has a left biadjoint.
44
STEPHEN LACK
Part (ii) is easier: we have a commutative diagram of 2-functors
T -Algs
J
fs∗
S-Algs
/ T -Alg
f∗
J
/ S-Alg
in which the left hand map has a left adjoint, by a general enriched-categorytheoretic fact (no harder than the corresponding fact for ordinary categories), and
the bottom map has a left adjoint (the pseudomorphism classifier for S-algebras).
Thus the composite has a left adjoint, and so f ∗ has a left biadjoint. (This argument, using the pseudomorphism classifier for S-algebras, requires S to have rank,
but this can be avoided.)
What about part (i)? For any S : C → T -Alg, we can form the diagram
J
/ T -Alg
OOO
OOO
OO
T -Alg(S,1)
T -Alg(S,J) OOO
'
Hom(C op , Cat)
T -Algs
and now the existence of bicolimits in T -Alg amounts to the existence of left biadjoints for all such T -Alg(S, 1). So it will suffice to show that the composite
T -Alg(S, J) : T -Algs → Hom(C op , Cat) has a left adjoint. But T -Alg(S, J) ∼
=
T -Algs (QS, 1), where Q a J, and T -Algs (QS, 1) has a left adjoint provided that
T -Algs has pseudocolimits. Finally since T is finitary, T -Algs is cocomplete (by a
general enriched-category-theoretic fact no harder than the corresponding fact for
ordinary categories) and so in particular has pseudocolimits.
A direct proof that T -Alg has bicolimits would be a nightmare, but using pseudocolimits it becomes manageable.
7. Model categories, 2-categories, and 2-monads
This section involves Quillen model categories, henceforth called model categories, or model structures on categories. There are various connections between
model categories, 2-categories, and 2-monads which I’ll discuss.
(i) Model structures on 2-categories: a model 2-category is a category with a
model structure and an enrichment over Cat, with suitable compatibility
conditions between these structures. Any 2-category with finite limits and
colimits has a “trivial” such structure, in which the weak equivalences are the
categorical equivalences. These trivial model 2-categories are not so interesting in themselves, but can be used to generate other more interesting model
2-categories.
(ii) Model categories for 2-categories: There’s a model structure on the category
of 2-categories and 2-functors, and one for bicategories too.
(iii) Model structures induced by 2-monads. If T is a 2-monad on a 2-category
K , we can lift the trivial model 2-category structure on K coming from the
2-category structure to get a model structure on T -Algs .
(iv) Model structures for 2-monads: the 2-category Mndf (K ) of finitary 2-monads
on K is also a model 2-category.
One thing which I won’t discuss, but deserves further study:
A 2-CATEGORIES COMPANION
45
(v) “Many-object monoidal model categories”. There’s a notion of monoidal
model category: this is a monoidal category with a model structure, suitably compatible with the tensor product. The many-object version of this
would involve a bicategory (or 2-category) with a model structure on each
hom-category, subject to certain conditions (somewhat more complicated than
those for monoidal model categories).
7.1. Model 2-categories. There’s a model structure on the category Cat0 of
categories and functors in which the weak equivalences are the equivalences of
categories, and the fibrations are the functors f : A → B such that for any object
a ∈ A and any isomorphism β : b ∼
=a
= f a in B, there is an isomorphism α : a0 ∼
0
in A with f a = b and f α = β. This is sometimes called the “categorical model
structure” or “folklore model structure”. (There are other model structures on
Cat0 , in particular the famous one due to Thomason that gives you a homotopy
theory equivalent to simplicial sets.)
As mentioned above, a category with a monoidal structure and a model structure
satisfying certain compatibility conditions is called a monoidal model category. The
cartesian product makes Cat0 into a monoidal model category.
If we now consider a category that has both a model structure and an enrichment
over Cat, there is a notion of compatibility between these structures, which can be
expressed in terms of the monoidal model structure on Cat0 . We call this notion
a model 2-category.
We start with a 2-category K with finite limits and colimits; equivalently, the
underlying category K0 of K has finite limits and colimits, and K has tensors and
cotensors with 2. Then K is a model 2-category if it is equipped with classes of
maps called the fibrations, the cofibrations, and the weak equivalences, satisfying
the usual model category axioms, plus a few new ones. These new axioms state
that for any cofibration i : A → B and fibration p : C → D:
(a) Given morphisms x : A → C, y : B → C, and z : B → D, with px = zi,
and invertible 2-cells α : x ∼
= yi and β : z ∼
= py with pα = βi, there exist a
0
morphism y : B → C and an isomorphism γ : y 0 ∼
= y with pγ = β and γi = α;
(b) If either i or p is trivial, then for any morphisms x, y : B → C and any 2-cells
α : xi → yi and β : px → py with βi = pα, there exists a unique 2-cell γ : x → y
with pγ = β and γi = α.
It follows that every equivalence is a weak equivalence, and that any morphism
isomorphic to a weak equivalence is itself a weak equivalence.
7.2. Trivial model 2-categories. Let K be a 2-category with finite limits and
colimits. The most important limit here will be the pseudolimit of an arrow f : A →
B. Ordinarily we don’t talk about limits of an arrow, since in an ordinary category
the limit of an arrow is just its domain, but the pseudolimit is only equivalent to
the domain, not equal. It’s the universal diagram
?A
~~
~
~~
~~
f
=λ
L @∼
@@
@@
v @@
B
u
46
STEPHEN LACK
such that given β : f a ∼
= b there is a unique c : X → L with λc = β (and so also
uc = a and vc = b). In this case, u is an equivalence, because id : f 1 ∼
= f factors
through by a d : A → L with ud = 1, and one can also check that du ∼
= 1. The
technique of Section 6.9 can be used to calculated the weight for pseudolimits of
arrows.
The model structure on K is:
• The weak equivalences are the equivalences;
• The fibrations are the isofibrations, the maps such that each invertible 2-cell
a
X
∼
=
b
)
A
f
0B
lifts to an invertible 2-cell
X
x
a
∼
=
a0
8A
f
0B
b
• The cofibrations have the left lifting property with respect to the trivial
fibrations.
• It follows that the trivial fibrations are the surjective equivalences: the p
for which there exists an s with ps = 1 and sp ∼
= 1.
We call such a model 2-category K a trivial model 2-category. When K = Cat
this is just the folklore structure. When K has no non-identity 2-cells, then the
equivalences are the isomorphisms, and all maps are isofibrations, so this agrees
with the usual notion of trivial model category.
The pseudolimit of f gives us, for any f , a factorization f = vd where v is a
fibration (which follows from the universal property of the pseudolimit) and d is
an equivalence. In the case of Cat, you could stop there and d would already
be a trivial cofibration, but in general there’s more work to do, although we have
reduced to the problem to factorizing an equivalence.
The way you do that is also the way you get the other factorization: use the dual
construction. Form the pseudocolimit of the arrow f , as in the diagram below, and
let e be the unique map with ei = f , ej = 1, and eϕ = id.
f
/C
~?
~
~
f
~~j
~
~
1
B
A
∼
=ϕ
i
e
'/
;B
This time i is a cofibration and e is a trivial fibration, and if f itself is an equivalence,
then i has the left lifting property with respect to the fibrations (so it’s what’s going
to become a trivial cofibration).
That’s all I’ll say about the proof. There is, of course, a dual model structure
in which the cofibrations are characterized and the fibrations are defined by a right
lifting property. For Cat, these coincide, in general they don’t.
A 2-CATEGORIES COMPANION
47
When K is arbitrary, there is no reason why the model structure should be
cofibrantly generated. Certainly for Cat it is, but even for such a simple 2-category
as Cat2 it is not. From the homotopical point of view the trivial model structure
is trivial in several ways, including:
• All objects are cofibrant and fibrant;
• The morphisms in the homotopy category Ho(K0 ) are the isomorphism
classes of 1-cells in K .
In the case K = Cat(E), one typically considers different model structures:
an internal functor F : A → B is usually called a weak equivalence if it’s full and
faithful and essentially surjective in an internal sense. For Cat this is equivalent
to the usual notion by the axiom of choice, but in general it won’t be. It’s the
weak equivalences in this sense that people tend to use as their weak equivalences
for Cat(E). When E is a topos, this was studied by Joyal and Tierney, and there’s
been recent work on other cases, when E is groups (so that internal categories are
crossed modules) or abelian groups.
7.3. Model structures for T -algebras. Now let T be a (finitary) 2-monad on
(a locally finitely presentable) 2-category K , and T -Algs the 2-category of strict
algebras and strict morphisms. In the usual way, one can lift the model structure
on K to get one on T -Algs : a strict T -morphism f : (A, a) → (B, b) is a weak
equivalence or fibration if and only if Us f : A → B is one in K ; the cofibration are
then defined via a left lifting property.
Now the lifted model 2-category structure on T -Algs is not trivial. In general,
if (f, f ) : (A, a) → (B, b) is a pseudomorphism of T -algebras and f : A → B
is an equivalence, then any inverse-equivalence g : B → A naturally becomes an
equivalence upstairs in T -Alg. This is a 2-categorical analogue of the fact that if
an algebra morphism is a bijection, its inverse also preserves the algebra structure.
But if f is strict (f is an identity), there is no reason why its inverse equivalence
should also be strict. For example, a strict monoidal functor which is an equivalence
of categories has an inverse which is strong monoidal, but which need not be strict.
Recall the adjoint to the inclusion
T -Algs
s
Q
⊥
3 T -Alg
J
where QA = A0 , so that we have a bijection
A
B
.
A0 → B
This fits into the model category framework very nicely. The counit of this adjunction
q
A0 −→ A
is a cofibrant replacement: a trivial fibration with A0 being cofibrant. So we see
that T -Alg, which is the thing we’re more interested in, is starting to come out of
the picture: a weak morphism out of A is the same thing as a strict morphism out of
the “special cofibrant replacement” A0 of A. This is much tighter than the general
philosophy that “we should think of maps in the homotopy category as maps out
of a cofibrant replacement.”
48
STEPHEN LACK
An algebra turns out to be cofibrant if and only if q : A0 → A has a section
in T -Algs . (There’s always a wiggly one.) Since q is a trivial fibration, there will
certainly be a section if A is cofibrant. Conversely, if there is a section, then A is a
retract of A0 ; but A0 is always cofibrant, and so then A must be cofibrant too. In
2-categorical algebra, the word flexible is used in place of cofibrant.
7.4. Model structures for 2-monads. Recall now that we have adjunctions
Mndf (K )
G
H
a W
H
a
Endf (K )
G
V
[obKf , K ]
both of which are monadic, as is the composite. Thus Mndf (K ) is both M -Algs
and N -Algs where M is the induced monad on Endf (K ) and N is the induced
monad on [obKf , K ].
Thus Mndf (K ) has two lifted model structures, coming from the trivial structures on Endf (K ) and on [obKf , K ]. They’re not the same, since something can
be an equivalence all the way downstairs without being one in Endf (K ) (which is
itself the 2-category of algebras for another induced monad on [obKf , K ]).
f
A monad map S −→ T is a 2-natural transformation compatible with the unit
and multiplication. If the 2-natural transformation is an equivalence in Endf (K ),
it is a weak equivalence for the M -model structure; if the components of the 2natural transformation are equivalences, it is a weak equivalence for the N -model
structure.
It’s the M -model structure (the one lifted from Endf (K )) which seems to be
more important, and we’ll only consider that one here. The corresponding prime
construction classifies pseudomorphisms of monads. These are precisely the things
that arise when talking about pseudoalgebras: recall that a pseudo-T -algebra was
an object A with a pseudo-morphism
T
hA, Ai
a
into the “endomorphism 2-monad” of A, corresponding to maps T A −→ A which
are associative and unital up to coherent isomorphism.
This corresponds to a strict map T 0 → hA, Ai, so that T 0 -Alg = Ps-T -Alg. (This
is the part of the justification for working with strict algebras that people tend to
understand first, but it’s the less important one: see Remark 4.1 above.)
If q : T 0 → T has a section in M -Algs = Mndf (K ), then T is said to be
flexible (=cofibrant). This was the context in which the notion of flexibility was
first introduced. Any monad that you can give a presentation for without
having to use equations between objects is always flexible. For example,
the monad for monoidal categories is flexible, but the monad for strict monoidal
categories is not.
Flexible monads have the property that every pseudo-algebra is (not just equivalent but) isomorphic to a strict one; in fact isomorphic via a pseudomorphism
whose underlying K -morphism is an identity! Remember that the importance of
A 2-CATEGORIES COMPANION
49
pseudo-algebras is not for describing concrete things, but for the theoretical side,
since various constructions don’t preserve strictness of algebras. For particular
structures like monoidal categories, you’re better off choosing the “right” monad
to start with: the one for which monoidal categories are the strict algebras.
7.5. Model structure on 2-Cat. 2-Cat is the category of 2-categories and 2functors. It underlies a 3-category, and a 2-category, and perhaps more importantly
a Gray-category. But we want to describe a model structure on the mere category
2-Cat, analogous to the one above for Cat.
The weak equivalences will be the biequivalences. Recall that F : A → B is a
biequivalence if
• each F : A (A, B) → B(F A, F B) is an equivalence of categories; and
• F is “biessentially surjective” on objects: if C ∈ B, there exists an A ∈ A
with F A ' C in B.
Every equivalence has an inverse-equivalence, going back the other way. For a
biequivalence F : A → B you can build a thing G : B → A with GF ' 1 and
F G ' 1. You can make G a pseudofunctor, but generally not a 2-functor, even
when F is one. That’s somehow the whole point of the model structure. Similarly
the equivalences F G ' 1 and GF ' 1 will generally only be pseudonatural.
Clearly biequivalence is the right notion of “sameness” for bicategories, or 2categories, but there is this stability (under biequivalence-inverses) problem, if you
want to work entirely within 2-Cat. If you allow pseudofunctors, and so move to
2-Catps , then as we have seen, you lose completeness and cocompleteness.
The fibrations are similar to the case of categories. Fibrations for the model
structure on Cat involved lifting invertible 2-cells; here we lift equivalences: a
2-functor F : A → B is a fibration if
• given an object A upstairs and equivalence downstairs, we have a lift as in
A_0
B
'
'
/A
_
A
B
F
/ FA
• given a 1-cell f upstairs and an invertible 2-cell downstairs, we have a lift
as in
*
=
5A
A
A_0 ∼
_
B
∼
=
(
6 FA
F
B
Equivalently, each of the functors A (A1 , A2 ) → B(F A1 , F A2 ) is an (iso)fibration
in Cat.
Note that the notion of biequivalence is not internal to the 3-category or Graycategory of 2-categories, 2-functors, and so on, which speaks against the existence
of a general model structure on an arbitrary 3-category or Gray-category which
would reduce to this one.
There’s an equivalent way of expressing these which is useful. Keep the iso-2cell lifting property as is, but modify the equivalence-lifting to deal with adjoint
50
STEPHEN LACK
equivalences rather than equivalences. Here it is not just the 1-cell, but also the
equivalence-inverse, and the invertible unit and counit which must be lifted.
In the presence of the iso-2-cell lifting property, these two types of equivalenceliftings are equivalent: clearly the lifting of adjoint equivalences implies the lifting
of equivalences, since we can complete any equivalence to an adjoint equivalence,
but the converse is also true provided that we can lift 2-cell isomorphisms.
This is related to a mistake I made in my first paper on this topic, where I used
a condition like this on lifting equivalences that aren’t necessarily adjoint equivalences. Regard “being an equivalence” as a property, and “an adjoint equivalence”
as a structure, but be wary of regarding “a not-necessarily-adjoint equivalence”
as a structure. Adjoint equivalences are now completely algebraic, classified by
maps out of “the free-living adjoint equivalence”, which is biequivalent to the terminal 2-category 1. A “free-living not-necessarily-adjoint equivalence” would not
be biequivalent to 1.
The trivial fibrations, which are the things which are both fibrations and weak
equivalences, can be characterized as the 2-functors that
• are surjective on objects; and
• have each A (A1 , A2 ) → B(F A1 , F A2 ) a surjective equivalence (a trivial
fibration in Cat).
Note that the trivial fibrations don’t use the 2-category structure; you don’t need
anything about the composition to know what these things are, only the “2-graph”
structure. So they’re much simpler to work with.
There’s an obvious ω-categorical analogue to these things, which permeates
Makkai’s work on ω-categories. You don’t need the ω-category structure, only
a globular set, to say what this means. The corresponding notion of “cofibrant
object” is then what he calls a “computad”.
It’s a bit less trivial than with the other model structures to prove that this all
works, but it’s not really hard. Everything is directly a lifting property (once you
use the version with adjoint equivalences), so finding generating cofibrations and
trivial cofibrations is easy.
All objects are fibrant, but not all objects are cofibrant. We have a “special”
cofibrant replacement q : A 0 → A with the property that that pseudofunctors out
of A are the same as 2-functors out of A ’:
A
B
A0 →B
and A is cofibrant (flexible) if and only if the trivial fibration q has a section in
2-Cat. This happens exactly when the underlying category A0 is free on some
graph (you haven’t imposed any equations on 1-cells, but you may have introduced
isomorphisms between them). In principle, a cofibrant A0 could be a retract of
something free, but it turns out that this already implies that it is free.
There are two main things of interest to me in this paper. The first was the
equation “cofibrant = flexible”. The second involved the monoidal structures. The
model structure is not compatible with the cartesian product ×. The thing to have
in mind is that the locally discrete 2-category 2 is cofibrant, but 2 × 2 is not, since
a commutative square
/
/
A 2-CATEGORIES COMPANION
51
is not free. There are various tensor products you can put on 2-Cat. The cartesian
product is also called the ordinary product (since it is also a special case of the
tensor product of V -categories), but I like to call it the black product since the
square is “filled in”.
There’s also the white or funny product, in which the square has nothing in it at
all. It’s a theorem that on Cat there are exactly 2 symmetric monoidal closed structures: the ordinary one and the “funny” one. The closed structure corresponding to
the funny product is the not-necessarily-natural transformations (just components).
Enriching over this structure gives you a “sesquicategory” (perhaps an unfortunate
name, but you can see how it came about), which has hom-categories and whiskering, but no middle-four interchange, hence no well-defined horizontal composition
of 2-cells. This funny tensor product can also be defined on 2-Cat, or indeed on
V -Cat for any V .
In the case of 2-Cat, there’s also the Gray or grey tensor product, due to John
Gray, in which you put an isomorphism in the square, so it’s “partially filled in”.
/
∼
=
/
(This is the “pseudo” version of the Gray tensor product; there’s also a “lax”
version: different shade of grey!)
The black and white tensor product make sense for any V at all, but the grey
one doesn’t. There’s a canonical comparison from the funny/white product to
the ordinary/black one, and the Gray/grey tensor product is a sort of “cofibrant
replacement” in between.
The Gray tensor product 2 ⊗ 2 is cofibrant, and more generally, the model
structure is compatible with the Gray tensor product.
Now consider Bicat, the category of bicategories and strict morphisms. Everything before looks exactly the same: the full inclusion 2-Cat ,→ Bicat preserves
and reflects weak equivalences and fibrations. This inclusion has a left adjoint “free
strictification”
t
2-Cat
⊥
4 Bicat .
It’s not a pseudomorphism classifier, since all morphisms here are strict, and it’s
not the usual strictification functor “st” either: in general the unit is not a biequivalence. But the component of the unit at a cofibrant bicategory is an equivalence.
This fits well into the model category picture: it’s part of what makes this adjunction a Quillen equivalence (one that induces an equivalence of homotopy categories).
There exist bicategories (even monoidal categories) for which there does not exist
a strict map into a 2-category which is a biequivalence, although we know that any
B has a pseudofunctor B
Bst . The point about cofibrant bicategories B is that
“there aren’t any equations between 1-cells”, and this is what makes the unit at
such a B an equivalence.
Just as for 2-Cat, we have pseudomorphism classifiers in Bicat, which serve as
“special cofibrant replacements”.
The model structure on 2-Cat is proper: showing that it’s left proper (biequivalences are stable under pushout along cofibrations) was the hardest part of that
52
STEPHEN LACK
paper, but the fact that it is right proper is an immediate consequence of the fact
that every object is fibrant.
7.6. Back to 2-monads. There’s a connection between the model structure on
Mndf (K ) and that on 2-Cat. There’s a 2-functor
sem : Mndf (Cat)op → 2-CAT/Cat
which you might call semantics, defined by:
U
T 7→ (T -Alg −→ Cat).
and
f
f∗
(S −→ T ) 7→ (T -Alg −→ S-Alg)
since if a : T A → A is a T -algebra, then its composite with f A : SA → T A makes
A into an S-algebra.
In the ordinary unenriched case or the V -enriched case, or even here, if we
used T 7→ T -Algs rather than T 7→ T -Alg, the semantics functor would be fully
faithful. But the semantics functor defined above, using T -Alg, is not: to give a map
sem(T ) → sem(S) in 2-CAT/Cat corresponds to giving a weak morphism from S
to T , but not in the sense of pseudomorphisms of monads, considered above; rather
in a still broader sense, in which the f A : SA → T A need not even be natural.
Now the definitions of fibration, weak equivalence, and trivial fibration in 2Cat have nothing to do with smallness, and make perfectly good sense in the
category 2-CAT of not-necessarily-small 2-categories. We can therefore define a
morphism in 2-CAT/Cat to be a fibration, weak equivalence, or trivial fibration
if the underlying 2-functor in 2-CAT is one.
Under these definitions, sem preserves limits, fibrations, and trivial fibrations, as
one verifies using the 2-monads hA, Ai, {f, f }` , and so on. Limits, fibrations, and
trivial fibrations in Mndf (Cat)op , correspond to colimits, cofibrations, and trivial
cofibrations in Mndf (Cat). Thus, it should in principle be the right adjoint part
of a Quillen adjunction. It’s not, of course, because of size problems: 2-CAT/Cat
has large hom-categories, and sem lacks a left adjoint.
The assertion that sem preserves the weak equivalence q : T 0 → T is equivalent
to the assertion that every pseudo T -algebra is equivalent to a strict one. More
generally, sem preserves all weak equivalences if and only if pseudo algebras are
equivalent to strict ones for every T . This is an open problem in the current
generality, but it is true that sem preserves weak equivalences between cofibrant
objects (flexible monads).
8. The formal theory of monads
In this section we return to formal category theory; in fact, to one of its high
points: the formal theory of monads.
8.1. Generalized algebras. Let’s start by thinking about ordinary monads. Let
A be a category, t = (t, µ, η) a monad on A. Write At for the Eilenberg-Moore category (the category of algebras). The starting point is to think about the universal
a
property of this construction. What is it to give a functor C −→ At ? We give for
each c ∈ C, an algebra ac, which we also use for the name of the underlying object,
A 2-CATEGORIES COMPANION
53
αc
with structure map tac −→ ac. And for every γ : c → d, we have an aγ : ac → ad
with a commutative square
αc /
ac
tac
aγ
taγ
tad
/ ad.
αd
This square looks an awful lot like a naturality square; it wants to say that α is
natural with respect to γ.
a
What we’re actually doing is giving a functor C −→ A and a 2-cell
/A
a
C
|
α
a
t
+A
with equations of natural transformations
/ ta o
µa
t2 a


α 
 1

/a
tα
ta
ηa
α
a
which just says that on components, it makes each ac into a t-algebra.
You might call this a generalized algebra, or a t-algebra with domain C. Think
of a usual algebra as a generalized algebra with domain 1.
Similarly, you can look at natural transformations. To give a natural transformation
a
C
'
t
7A
b
amounts to giving
a
C
ϕ
%
9A
b
which is suitably compatible, in the sense that
ta
tϕ
α
/ tb
β
a
ϕ
/ b.
This is the universal property of the Eilenberg-Moore construction, and the starting
point of the theory.
I’ve been talking all along about categories, but once we’ve moved beyond algebras with domain 1, there’s no reason to restrict in that way, so we can talk instead
about a monad on an object A in any 2-category K . (The notion of monad has
not been weakened in any way. The 2-category K might be Cat, or 2-Cat, or
V -Cat, but we use the same definition.)
54
STEPHEN LACK
We can’t just construct At as we did before, but we can ask whether there exists
an object At with the universal property. A slick way to do this is as follows. The
hom-category K (C, A) has a monad K (C, t) on it (since 2-functors take monads
to monads), and this is the ordinary type of monad in Cat. The endofunctor part
of this monad sends a : C → A to ta : C → A. This generalized notion of algebra
is then nothing but the usual sort of algebra for the ordinary monad K (C, t). So
what we want is an isomorphism
K (C, At ) ∼
= K (C, A)K (C,t)
naturally in C (where the right hand side means the ordinary Eilenberg-Moore
category of algebras for the ordinary monad K (C, t)). We call At the EilenbergMoore object of t, or EM-object for short.
It turns out that in some places, such as Cat, it’s enough to check that the
universal property for C = 1, but in an abstract 2-category there may not be a 1,
and if there is, it may not be enough to get the full universal property.
The universal property of At makes it look like a limit, and indeed it is, but we’ll
look at some other points of view first.
8.2. Monads in K . Let K be a 2-category. Previously we looked at the 2category Mndf (K ) of finitary 2-monads on K (as a fixed base object). We now
consider the 2-category mnd(K ) of all the (internal) monads in K , with variable
base object.
• Its objects are monads in K .
• Its 1-cells correspond to morphisms which lift to the level of algebras:
m
At
/ Bt
us
ut
A
m
/B
and we can think of this as an identity 2-cell and take its mate, since the
the us are right adjoints:
m
A
} ft
At
m
/B
fs
/ Bt
us
ut
A
m
/B
which we then paste together to get a 2-cell, and the forgetful-free composite
gives us the monads. Thus we should define a morphism of monads to be
a morphism m : A → B with a 2-cell ϕ : sm → mt such that the diagrams
ssm
sϕ
/ smt
µm
sm
ϕt
/ mtt
mµ
ϕ
/ mt
{
{{
{
{{
{} {
sm
ηm
mC
CC
CCmη
CC
C!
/ mt
λ
A 2-CATEGORIES COMPANION
55
commute. We could define a morphism of monads to be a morphism m :
A → B with a lifting m : At → B s as above, except that the EM-objects
need not exist in a general 2-category; but when they do exist, the two
descriptions are equivalent.
• The 2-cells in mnd(K ) are 2-cells
m
A
ρ
&
8B
n
in K with a compatibility condition, which you could express as saying
that ρ lifts to a ρ between EM-objects, or saying that
ϕ
sm
/ mt
sρ
ρt
sn
ψ
/ nt
commutes.
There’s a full embedding id : K ,→ mnd(K ) sending A to the identity monad
(A, 1) on A, and the obvious thing for 1-cells and 2-cells. This is particularly clear
in the EM-objects picture, since if t = 1 then At = A, so obviously m will lift
uniquely to an m, which is what fully faithfulness of id says.
A trivial observation is that for any monad we can always choose to forget the
monad and be left with the object, and this is left adjoint to id. The more interesting
thing, however, is the existence of a right adjoint: this amounts exactly to a choice
of an EM-object for each monad in K . Why? Look at the universal property: If
A 7→ At is the right adjoint, this says that
K (C, At ) ∼
= mnd(K )((C, 1), (A, t))
The key point is that the right hand side is equal to K (C, A)K (C,t) , since we get
a
C
1
C
|
α
a
/A
t
/A
with exactly the conditions which make (a, α) into a generalized algebra.
Now the really beautiful thing happens: we can start looking at duals of K and
see what happens. Consider first K co , where we reverse the 2-cells but not the
1-cells. A monad in K co is then a comonad in K . And an EM-object in K co is
the obvious analogue for comonads. If K = Cat, we get ordinary comonads, and
the EM-object is the usual category of coalgebras for the comonad.
That’s nice, but not incredibly surprising. What’s more interesting is what
happens in K op . A monad in K op consists of an object A, a morphism not from
A to A, but rather (!) from A to A:
Ao
t
A
and for the multiplication you have to make sure when you compose t with itself,
you do it in the reverse order, and so on. But all this just amounts to . . . a monad
in K .
56
STEPHEN LACK
But what about the EM-object? The arrows are reversed, so we get a different
universal property. An algebra for this monad consists of
CS o
a
|
AO
α
t
a
A
The amazing thing is that in the case K = Cat this is the same thing as a map
At → C where At is the Kleisli object. Therefore Eilenberg-Moore objects in K co
are called Kleisli objects (in K ).
It’s true in any 2-category that the EM-object is the terminal adjunction giving
rise to the monad, and the Kleisli object is the initial one, but the universal property
given above is richer in that it refers to maps with arbitrary domains.
Using K coop , of course, gives you Kleisli objects for comonads.
8.3. The monad structure of mnd. Now, where does the construction mnd(K )
really live? Consider the category 2-Cat of 2-categories and 2-functors. Completely
banish from your mind all concerns about size, which doesn’t have any role here.
So far we’ve constructed a 2-category mnd(K ) for any 2-category K . and this is
clearly completely functorial, so we get a functor
mnd : 2-Cat → 2-Cat
and the inclusion id is clearly natural in K , so we get a natural transformation
1
2-Cat
*
id
42-Cat
mnd
A certain sort of person is tempted to wonder whether this is part of the structure
of a monad on 2-Cat! We do have a composition
mnd2
2-Cat
comp
,
2 2-Cat
mnd
and what it does is the other beautiful thing about the paper.
This composition map sends a monad in mnd(K ) to a monad in K . What is
a monad in mnd(K )? It consists of
• a monad (A, t) in K (an object of mnd(K ))
s
• an endo-1-cell, which consists of a morphism A −→ A in K with a 2-cell
λ
ts −→ st (with conditions)
ν
• A multiplication (s, λ)(s, λ) → (s, λ), corresponding to s2 −→ s (with conditions)
• a unit 1 → (s, λ) corresponding to 1 → s (with conditions)
As well as the conditions for these to be 1-cells and 2-cells in mnd(K ), we need
the conditions for this to be a monad there. These make s itself into a monad on
A in K . The 2-cell λ is now what’s called a distributive law between these two
monads, which is exactly what you need to “compose” these two monads and get
another monad.
A 2-CATEGORIES COMPANION
57
Think about this as like the tensor product of rings. R ⊗ S is the tensor product
of the underlying abelian groups, with multiplication
R⊗S⊗R⊗S
1⊗tw⊗1
/ R ⊗ R ⊗ S ⊗ SmR ⊗mS
/ R ⊗ S.
The point is we’re trying to do something very similar, but here we’re in a world
where the tensor product is not commutative, so we don’t have the twist. So λ plays
the role of the twist; it’s a “local” commutativity that only applies to these two
objects. The conditions put on it are exactly what we need to make the composite
st into a monad.
For example, the multiplication on st is then
sλt
ssµ
µt
stst −→ sstt −→ sst −→ st
The notion of distributive law, in the ordinary case of categories, is due to Jon
Beck, and he proved that we have a bijection between distributive laws ts → st and
“compatible” monad structures on st, and also to liftings of s to At (whenever At
exists). It’s not as well known as it should be and is frequently rediscovered.
You can also do this for K co or K op or K coop , of course. A distributive law in
K op is formally the same as a distributive law in K , but now rather than liftings
of s to At , one has extensions of t to As (along the left adjoint fs : A → As .
Remark 8.1. Operads are monoids in a monoidal category, so there is a corresponding notion of distributive law between operads. Furthermore, the passage from the
monoidal category of collections to the monoidal category of endofunctors is strong
monoidal, so distributive laws between operads induce distributive laws between
the induced monads, and this process is compatible with the formation of the composite operad/monad. Just as not every monad arises from an operad, not every
distributive law between monads arises from a distributive law between operads,
even when the monads themselves do arise from operads.
Example 8.2. Groups are particular monoids in Set, so there is a corresponding
notion of distributive law. If a group G acts on a group H, then there is a distributive law G × H → H × G sending (g, h) to (g.h, g), and the induced “composite” is
the semidirect product H o G. This generalizes to arbitrary monoids in a cartesian
monoidal category.
8.4. Eilenberg-Moore objects as limits. There are two ways to see EilenbergMoore objects as weighted limits. Remember that way back in the first week, we
saw that monads t in K correspond to lax functors t̃ : 1 → K . Then the lax limit
of t̃ is exactly the EM-object At .
I didn’t explicitly talk about lax limits of lax functors, but it’s not hard to extend
the definition of lax limit to cover this case. Alternatively one can replace the lax
functor by the corresponding 2-functor out of the “lax morphism classifier”, and
then just take the lax limit of the 2-functor. Let’s see how this would work.
First recall how t̃ is defined. It sends ∗ to A, and 1∗ to an endomorphism
t : A → A, the unit is the lax unit comparison, and the multiplication is the lax
composition comparison. To understand the lax limit of these sorts of things, we
should think about lax cones. A lax cone would involve a vertex C of K , with just
58
STEPHEN LACK
a
one component C −→ A, and a lax naturality 2-cell for every 1-cell in 1:
/A
a
C
|
C
α
a
t
/A
and some conditions.
There’s an old paper of Street called “two constructions on lax functors”, and
this is the first construction. The second was the lax colimit, which gives the Kleisli
construction.
The lax morphism classifier on 1 is a 2-category mnd with a bijection
2-functors mnd −→ K
lax functors 1 −→ K
but such lax functors are in turn the same as monads in K . Thus mnd is the
universal 2-category containing a monad. Remember that a monad in K is the
same as a monoid in a hom-category, and we know the universal monoidal category
containing a monoid is the “algebraic ∆”, the category Ordf of (possibly empty)
finite ordinals. This is not the ∆ of simplicial sets: an extra object has been added.
Thus mnd has one object ∗ and mnd(∗, ∗) = Ordf .
Now we have a limit notion (t̃ 7→ At ), and we want to know the corresponding
weight J : mnd → Cat, so that {J, t̃} = At . We saw in Section 6.9 that the recipe
for calculating J is to consider the Yoneda functor mnd → [mnd, Cat]op and form
the limit of it, or equivalently the colimit of mndop → [mnd, Cat]. The colimit is
the Kleisli object; since we are in a presheaf 2-category [mnd, Cat] it is computed
pointwise. The weight is called alg; it’s now a straightforward exercise to calculate
it.
Of course, in general, alg-weighted limits may or may not exist. Subject to the
existence of the relevant limits, they can be built up from other limits we already
know:
t
• First form the inserter of A
&
k
8 A . This is an A1 −→ A equipped with
1
κ
a 2-cell tk −→ k.
k0
• Then take the equifier of k(ηk) and 1 to get an A2 −→ A1 such that the
identity law holds.
• Finally take the equifier of something else to get the associativity.
In particular, this shows that EM-objects are PIE-limits, in fact finite PIE-limits.
8.5. Limits in T -Alg` and T -Algc . T -Alg` and T -Algc are, recall, the 2-categories
of strict T -algebras with lax and with colax morphisms. Recall also that we had
nice pseudo-limits in T -Alg; here it’s much harder.
In T -Alg` , you have oplax limits, and in T -Algc you have lax limits (it’s a twisted
world we live in!) These are much more restricted classes of limits, not including
inserters, equifiers, comma objects and many of our other favourite limits.
I described how to construct inserters and equifiers in T -Alg: form the limit
downstairs and show that the thing you get canonically becomes an algebra. This
involves morphisms (f, f ) and (g, g) between T -algebras (A, a) and (B, b), and 2cells f k → gk for some k. If you look carefully at the construction, you’ll see that
A 2-CATEGORIES COMPANION
59
f needs to be invertible, but g can be arbitrary. So you can form inserters and
equifiers in T -Alg` provided that one of the 1-cells (the one that is, or tries to be,
the domain of the 2-cells) is actually pseudo.
Dually, in T -Algc , it’s the other 1-cell which needs to be pseudo. Now the
Eilenberg-Moore object of a monad (A, t) can be constructed using the inserter
k : C → A of t and 1A , and then an equifier (see Section 8.4). Furthermore 1A will
always be strict, and it turns out that T -Algc does have Eilenberg-Moore objects
for monads. The most important case is where T -algebras are monoidal categories
and so T -Algc has opmonoidal functors. Then a monad in T -Algs is an ordinary
monad for which the category is monoidal, the endofunctor opmonoidal, and the
natural transformations are opmonoidal natural transformations; this is sometimes
called a Hopf monad.
8.6. FTM 2. We can now see EM-objects as weighted limits in the strict sense,
and there’s a well-developed theory of free completions under classes of weighted
limits. So we can form the free completion EM(K ) of a 2-category K under
EM-objects; or we can form the corresponding colimit completion KL(K ), which
freely adds Kleisli objects. These are related: EM(K ) = KL(K op )op .
The colimit side is more familiar to construct. To freely add all colimits to an
ordinary category, we take the presheaf category; to add a restricted class, we take
the closure in the presheaf category under the colimits we want to add. So here, to
get KL(K ), we take the closure of the representables in [K op , Cat] under Kleisli
objects. It’s part of a general theorem that this works, at least when K is small.
Sometimes it can be tricky to calculate exactly which things appear in this
completion process. You start with the representables and chuck in the Kleisli
object for any monad. Usually this is an iterative process, since there will be new
diagrams appearing at each step and you have to continue, possibly transfinitely.
The nice thing about this particular case is that, as we shall see, it stops after one
step.
Colimits in the functor category are constructed pointwise, so we construct Kleisli
objects as in Cat. The key facts are:
• The Kleisli adjunctions in Cat are precisely the bijective-on-objects left
adjoints.
• These are closed under composition.
Now, given a monad t on A we throw in the Kleisli object At in [K op , Cat], which
may have a new monad s on it. We then throw in its Kleisli object for s to get
(At )s , but then the composite
A −→ At −→ (At )s
is also a bijective-on-objects left adjoint, hence (At )s is also a Kleisli object for a
monad on A. Thus this is a 1-step process.
Therefore, we can identify (up to equivalence) the objects of KL(K ) with monads in K , and then explicitly describe morphisms and 2-cells between them in
terms of K itself.
In the dual case EM(K ) = (KL(K op ))op we get
• The objects are the monads in K ,
• The morphisms are the monad morphisms (same as in mnd(K )), and
60
STEPHEN LACK
(m,ϕ)
• The 2-cells (A, t)
(n,ψ)
+
3 (B, s) are 2-cells m → sn (which should look
“Kleisli-like”) with some compatibility with t.
Composition is a also Kleisli sort of thing. Think of sn as the “free s-algebra on
n”, so using the universal property of free algebras, can express this as something
sm → sn, and express compatibility that way.
Why is this a good thing to do?
(i) We still have a fully faithful inclusion id : K → EM(K ), and a right adjoint
to this is just, by general nonsense for limit-completions, to give a choice of
EM-objects in K .
(ii) This comes up in examples. If we start with Span, we’ve seen that categories
are just monads in Span, and that functors can be seen as special morphisms
between such monads; now we can also deal with natural transformations.
There is a 2-functor
Cat → KL(Span)
which is bijective-on-objects and locally fully faithful, so that KL(Span) captures precisely the notion of natural transformation. This works equally well
for Cat(E), for V -Cat, or for generalized multicategories.
(iii) Remember that a distributive law is a monad in mnd(K ). The multiplication
and unit are 2-cells in mndK , so if we change the 2-cells, the notion of monad
changes. A monad in EM(K ) is more general: we call it a wreath, since the
composition operation is a wreath product.
A wreath still lives on a monad (A, t) in K . We have an endomorphism s :
A → A as before, along with a λ : ts → st with some conditions as before, but s
is no longer a monad: the multiplication is now something ν : ss → st, and the
unit σ : 1 → st. You can still make sense of associativity and unit using λ, but
everything ends up in st. Ultimately this gives a monad structure on st, which is
called the wreath product or composite of s and t.
For example, consider the monoidal category Set under cartesian product. This
can be regarded as a one-object bicategory, and so, after strictification, as a 2category. Let G be a group acting on an abelian group A, and consider a normalized
ρ
2-cocycle G × G −→ A. We consider A and G as monoids, hence monoids in Set.
A is our monoid. G happens also to be a monoid (in fact a group), but the monoid
structure isn’t used directly. We have the action
λ: G × A → A × G
(g, a) 7→ (g a, g)
and our
ν : G × G −→ A × G
(g, h) 7→ (ρ(g, h), gh)
this is a wreath, so it induces a monoid structure A o G (which is actually a group).
The multiplication is the usual one coming from the cocycle.
There’s a corresponding thing for Hopf algebras, giving a type of “twisted smash
product”, although not the most general.
A 2-CATEGORIES COMPANION
61
8.7. Another point of view on EM(K ). Here’s another point of view. It’s particularly suggestive if we take K to be the monoidal category (1-object bicategory)
Ab of abelian groups. Then a monad (monoid) in Ab is a ring R: the objects of
EM(Ab) are the rings.
We defined a morphism (f, ϕ) : (A, t) → (B, s) in EM(K ) to consist of a 1-cell
f : A → B and a 2-cell ϕ : sf → f t subject to two equations. A 1-cell R → S in
EM(Ab) consists of an abelian group M and a map S ⊗ M → M ⊗ R. Think of
this as being a bimodule structure on M ⊗ R; the left action is
S ⊗ M ⊗ R −→ M ⊗ R ⊗ R −→ M ⊗ R
and the right action is the free one, and the conditions on ϕ make it work. Thus
the 1-cells are the right-free bimodules. The 2-cells are then just module maps.
Composition of 1-cells is the ordinary module composition, but because of the
freeness condition, don’t need to use any coequalizers. If we were to look at KL(K ),
we’d get the left-free modules.
One could also consider arbitrary modules. This is an important construction,
but it requires the bicategory to have coequalizers in the hom-categories in order to
define composition; and these coequalizers to be preserved by whiskering on either
side in order for this composition to be associative (up to isomorphism), and so this
has rather a different flavour.
9. Pseudomonads
These are formally very similar to monoidal categories. A pseudomonad involves
a thing T , which plays the role of a category, a multiplication m : T 2 → T , a unit
i : 1 → T , an associativity isomorphism
/ T2
T3
T2
∼
=
/T
unit isomorphisms λ, ρ, and so on, all looking very like a monoidal category.
Just as monads can be defined in any 2-category or bicategory, pseudomonads can
be defined in any Gray-category or tricategory. The monoidal 2-category Cat (with
cartesian structure) can be regarded as a one-object tricategory, and a pseudomonad
in this tricategory is precisely a monoidal category. The associativity pentagon
becomes a cube, relating ways to go from T 4 to T , involving a bunch of µ’s and
a pseudonaturality isomorphism. In monoidal categories, one side of the cube
corresponds to
((A ⊗ B) ⊗ C) ⊗ D
A ⊗ (B ⊗ (C ⊗ D))
(A ⊗ (B ⊗ C)) ⊗ D
TTTT
5
TTTT
jjjj
j
j
j
TTTT
j
jjj
TTT)
jjjj
A ⊗ ((B ⊗ C) ⊗ D)
62
STEPHEN LACK
and the other side corresponds to
(A ⊗ B) ⊗ (C ⊗ D)
5
TTTT
jjjj
TTTTT
j
j
j
j
TTTTT
jj
j
j
j
T
jj
((A ⊗ B) ⊗ C) ⊗ D
(A ⊗ B) ⊗ (C ⊗ D)
A ⊗ (B ⊗ (C ⊗ D))
while in general, the equality will be replaced by an isomorphism, saying that it
doesn’t matter whether we tensor A and B first, then C and D, or vice versa.
Our unit isomorphisms will be
/ T2
T B
BB BB}
=E BB
B /T
T2
we could take them going the same way, to avoid using inverses (say, if we cared
about lax things), but this way will be convenient for the coherence result.
It is convenient to work with Gray-categories rather than tricategories; by the
coherence result that every tricategory is triequivalent to a Gray-category there is
no loss of generality. Note, however, that the one-object tricategory corresponding
to Cat is not a Gray-category, although it is a very special sort of tricategory.
One reason for working with Gray-categories is that we can then make use of
the huge amount of machinery developed for enriched categories.
9.1. Coherence. The coherence result describes the fact that there’s a universal
Gray-category with a pseudomonad in it: there’s a Gray-category Psm such
that for any Gray-category A, to give a Gray-functor Psm → A is equivalent to
giving a pseudomonad in A.
Psm is sort of a cofibrant resolution of mnd. More precisely, Psm like mnd has
a single object ∗, and Psm(∗, ∗) is a cofibrant replacement of mnd(∗, ∗). It’s not
a pseudomorphism classifier: that would be too large; we need a smaller cofibrant
resolution. Recall that mnd(∗, ∗) = Ordf = ∆, the category of finite ordinals,
or “algebraists’ simplicial category”. The underlying category of Psm(∗, ∗) (which
is a 2-category, since Psm is a Gray-category) is freely generated by the face and
degeneracy maps in ∆ (forget the relations we expect to hold)
/
/ o
...
/
Since this graph G generates ∆, we have a map F G → ∆ which is bijective on
objects and surjective on objects, so we can factor it as a b(ijective on) o(bjects)
b(ijective on) a(rrows) 2-functor followed by an l(ocally) f(ully) f(aithful) one (throw
in isomorphisms between the things that would become equal in ∆), to get
/
FG J
u: ∆
JJ
u
u
JJ
uu
JJ
uu
JJ
u
$
u
Psm(∗, ∗)
A 2-CATEGORIES COMPANION
63
To construct ∆0 , we would forgot all the way down to the underlying graph of ∆,
rather than a generating graph for it, and that would produce all sorts of stuff that
we don’t really need. It would include, for example, a generating operation T n → T
for any n
We also saw something like this for the Gray tensor product, factoring the map
from the funny tensor to the ordinary one.
You now have to define the composition in Psm
Psm(∗, ∗) ⊗ Psm(∗, ∗) −→ Psm(∗, ∗)
to make it a Gray-category. You basically take the composition in ∆, use that to
define it on the generators, then build it up to deal with arbitrary 1-cells, but since
the relations only hold up to isomorphism, that’s why the Gray-tensor appears.
Now you prove that this has the universal property that I said it does, so it
really does classify pseudo-monads in a Gray-category. I’m certainly not going to
do that. Roughly, how does it go? Given a pseudomonad, we have
iT
i
/T o
/
m
...
T2
/
and so on, which defines the putative Gray-functor Psm → A on objects, 1-cells,
and 2-cells. The fun starts when we come to the 3-cells: we have µ, λ, and ρ, and
we need to build up all the other required 3-cells. The idea is that for any 2-cell f
in Psm (any 1-cell in the above picture, generated by m’s and i’s), there is a normal
form f and a unique isomorphism f ∼
= f built up out of the 3-cells in Psm that
one might expect to call µ, λ, and ρ. Thus any 3-cell f ∼
= g in Psm can be written
as a composite f ∼
=f =g ∼
= g, and this can be used to define the Gray-functor
Psm → A on a 3-cells. The details of the rewrite system that these normal forms
come from are a bit technical.
1
Ti
9.2. Algebras. The next step is to construct a particular weight Psa : Psm →
Gray such that for any Gray-functor T : Psm → A, the weighted limit {Psa, T}
is the object of pseudoalgebras, pseudomorphisms, and algebra 2-cells (all suitably
defined) for the pseudomonad corresponding to T. Again, this is sort of a “cofibrant
replacement” for the corresponding one for 2-categories, although the domain has
changed.
I won’t do this, but I do want to make one point. It is the fact that we are working
with Gray-categories rather than 3-categories which causes the pseudomorphisms
to appear here. Recall that for ordinary monads, we talked about the fact that to
give something C → At is the same as a : C → A with an action α : ta → a, where
c 7→ (αc : tac → ac), and γ : c → d is sent to
tac
αc
aγ
taγ
tad
/ ac
αd
/ ad
and the fact that aγ is a homomorphism can be seen as the naturality of α. There’s
an analogous fact for operads and Lawvere theories: the actions are natural with
respect to homomorphisms.
64
STEPHEN LACK
When we come up to the Gray situation, we are thinking of pseudonatural
transformations, hence the square commutes up to isomorphism, so we get pseudomorphisms, not strict ones. That’s the “reason” for making the formal theory
of pseudo-monads live in the Gray context. Even if you wanted only to consider
3-categories A, the fact of working over Gray gives you the pseudomorphisms.
10. Nerves
In this section we use ∆ for the “topologists’ delta”, the category of non-empty
finite ordinals. As usual, we write [n] for the ordinal {0 < 1 < . . . < n}. This
section is particularly light on details; see [33] for more.
The nerve of an ordinary category C is the simplicial set N C in which
•
•
•
•
a 0-simplex is an object
a 1-simplex is a morphism
a 2-simplex is a composable pair and its composite
and so on.
This process gives a fully faithful embedding
Cat ,→ [∆op , Set]
.
The nerve of a bicategory B is the simplicial set N B in which
• a 0-simplex is an object
• a 1-simplex is a morphism
• a 2-simplex is a 2-cell living in a triangle
? ??
 ⇓ ???

??


/
• and so on.
These 2-simplices are being overworked; they have to express at the same time
composition of 1-cells, at least in some weak way, and what the 2-cells are. The
problem is that they don’t really ever say what the composite of a 1-cell is, only
what maps out of it are. Now that has its advantages, but it does make it hard
to say when such composites are being preserved. In fact we get a fully faithful
embedding
Bicatnlax ,→ [∆op , Set]
where “nlax” is short for “normal and lax”, which means morphisms which are
strict with respect to identities, but lax with respect to composition.
A lot of the time you want to talk about homomorphisms rather than lax ones.
If you want to get your hands on those there are various possibilities. One is to have
a bit more structure than a simplicial set: specify as extra data which 2-simplices
actually have an equality, or an isomorphism (and similar for higher simplices). A
simplicial set equipped with such a chosen class of simplices is called a stratified
simplicial set. One can characterize the stratified simplicial sets which are stratified
nerves of 2-categories or (via a different stratification) of bicategories; and indeed
similarly for strict or weak ω-categories. The stratified simplicial sets of the latter
characterization are called complicial sets.
A 2-CATEGORIES COMPANION
65
A different way to specify this extra structure is to use simplicial objects not in
Set but in Cat. For a bicategory B, the 2-nerve N2 B of B (or just N B from now
on) is a functor N B : ∆op → Cat
• (N B)0 is the discrete category of the objects.
• (N B)1 is category whose objects are morphisms and whose morphisms are
2-cells
> . So far it looks like some kind of enriched nerve.
• (N B)2 doesn’t need to include the 2-cells, since we already have them; we
can therefore take the objects of (N B)2 to be the isomorphisms
? ???
?
 ∼

 = ???


/
and a morphism of (N B)2 to consist of three 2-cells which commute in the
obvious way (in particular, the objects are all fixed).
• an object of (N B)3 is a tetrahedron all of whose faces are isomorphisms,
and so on.
We’d like a nice functorial description of the 2-nerve. Consider NHom, the
2-category of bicategories, normal homomorphisms, and icons (which, recall, are
oplax natural transformations all of whose 1-cell components are identities). Now
Cat ,→ NHom, where Cat is the locally discrete 2-category consisting of categories, functors, and only identity natural transformations, embedding as a full
sub-2-category consisting of the locally discrete bicategories. (An icon between
functors can only be an identity.) And of course we have ∆ ,→ Cat, so the composite fully faithful J : ∆ ,→ NHom induces
NHom(J, 1) : NHom −→ [∆op , Cat]
sending B to NHom(J−, B).
For instance, [0] ∈ ∆ goes to the terminal bicategory, and a normal homomorphism from that into B is just an object of B, with no room for icons; [1] ∈ ∆
goes to the arrow category 2, so a normal homomorphism from this into B is an
arrow, and the icons are exactly what we want.
Theorem 10.1. NHom(J, 1) = N is a fully faithful 2-functor (in a completely
strict sense) and has a left biadjoint (following from 2-categorical nonsense).
How can we characterize the image? X ∈ [∆op , Cat] is isomorphic to some N B
if and only if
(a) X0 is discrete.
(b) X is 3-coskeletal ; that is, isomorphic to the right Kan extension of its 3truncation — the idea is that 4-simplices and higher are uniquely determined
by their boundary.
(c) X2 → cosk1 (X)2 is a discrete isofibration. A functor p : A → B is a discrete
isofibration if given e ∈ E and β : b ∼
= pe, there exists a unique ε : e0 ∼
= e with
pε = β. This implies that if
X
ε
&
8E
66
STEPHEN LACK
and pε = id, then ε = id.
(d) X3 → cosk1 (X)3 (could also use the 2-coskeleton) is also a discrete isofibration
(e) The Segal maps are equivalences.
A Tamsamani weak 2-category, or just Tamsamani 2-category, since no strict
notion is considered, is a functor X : ∆op → Cat satisfying (a) and (d); thus the
2-nerve of a bicategory is a Tamsamani 2-category. Tamsamani suggests a way
of getting from a bicategory to a Tamsamani 2-category, but it is not the 2-nerve
construction given here.
The inclusion of NHom into Tamsamani 2-categories looks like it should be a
biequivalence, but it’s not quite. It would be if you broadened the definition of
morphism of Tamsamani 2-category to include what might be called the “normal
pseudonatural transformations”
What you might guess for the nerve of a bicategory is to have
•
•
•
•
N B0
N B1
N B2
N B3
thePobjects
= x,y B(x, y)
P
= x,y,z B(y, z) × B(x, y)
P
= w,x,y,z B(y, z) × B(x, y) × B(w, x)
which is what we do for the Cat case. If you try to do this, the simplicial identities
fail, due to the failure of associativity. Actually, what we do is to take the pseudolimit of the composition functor
X
X
B(y, z) × B(x, y) −→
B(x, y).
x,y,z
x,y
And this goes on; for composable triples, we have
/ B2
B3
B2
∼
=
/B
and N B3 is the pseudo-limit of this whole diagram. Going on, we can take the
pseudo-limit of all sorts of various higher cubes. It’s actually even true at N B1 , if
you say what you mean, but not very helpful.
11. Commented bibliography
The basic references for bicategories/2-categories are [2], [13], [24], and [50]. The
basic references for enriched categories are [11], [21], and [36]. For a good example
of simplicially-enriched category theory that is very close to 2-category theory, see
[9]. Both 2-categories and double categories were first defined by Ehresmann (see
perhaps [10]); bicategories were first defined by Bénabou [2]. For (a generalization
of) the fact every bicategory is biequivalent to a 2-category, see [37].
For monads in 2-categories see the classic [46]. For extensions and liftings see
[55]. For the calculus of mates, see [24], and for doctrinal adjunction see [17].
The importance of monoidal functors (not necessarily strong) was observed both
by Eilenberg-Kelly [11] and by Bénabou [2]. The importance of lax functors, especially lax functors with domain 1, was observed by Bénabou [2].
A 2-CATEGORIES COMPANION
67
Categories enriched in a bicategory were first defined by Walters to deal with
the example of sheaves on a site [59, 60]. A good general reference is [3]. Twosided enrichments (although not from the point of view of partial morphisms of
bicategories) were defined in [25].
The basic reference for 2-monads is [5], although many of the basic ideas go all
the way back to [16], including the constructions hA, Ai and {f, f } which allow one
to describe algebras in terms of monad morphisms. For the latter, see also [22]. The
accessibility issues in [5] were treated in Blackwell’s (unpublished) thesis, and later
in the monumental (and somewhat impenetrable) [18]. Anyway, [5] contains the
results about limits and (bi)colimits in T-Alg, biadjoints to algebraic 2-functors,
and the left adjoint to T-Algs → T-Alg. The proof I gave of the existence of this
left adjoint came from [28]. For locally finitely presentable categories see [12] or
[1]; and for the enriched version see [19]. For presentations for 2-monads see [23]
(and also [26] for the extra monadicity result). The idea of showing that pseudoT -algebras are equivalent to strict ones came up in question time: see [28] for a
general overview and more references, but I must mention here also the short and
beautiful paper [43] of Power.
Many people came up with some notion of weighted limit at about the same
time. But I guess the main reference for general V is now just [21]. On the other
hand, for various limit notions for 2-categories, [20] is very readable. Once you’ve
got through that, you should turn to [4]. For the beautiful theory of PIE-limits,
see [45]. For the connection between pullbacks and pseudopullbacks, see [14].
The “categorical” model structure on Cat seems to be folklore; the first reference
I know is [15]. The Thomason model structure [57] on Cat has an error in the proof
of properness which was corrected in [8]. For the model structures on 2-Cat and
Bicat see [29, 30]. There is also a “Thomason-style” model structure on 2-Cat
due to Hess, Parent, Tonks, and Worytkiewicz [62]. The model structure on the
category of monads, and its relation to structure and semantics, is mine [32].
The formal theory of monads goes back of course to [46]; for the account using
limit-completions, and the notion of wreath see the much later sequel [34]. The
Eilenberg-Moore object was described as a lax limit of a lax functor in [47], and
as a weighted limit in [48]. For limits in T-Algl , including Eilenberg-Moore objects
for comonads, see [31]; for Hopf monads see [41] and also [40].
The basic definitions involving pseudomonads in Gray-categories were given in
[39, 38]; for the universal pseudomonad, and the Gray-limit approach to pseudoalgebras, see [27].
The notion of nerve of a bicategory is due to Street. For nerves of ω-categories as
stratified simplicial sets, see [58], and the references therein. The notion of 2-nerve
of a bicategory is described in my paper with Paoli [33]. Tamsamani’s definition of
weak n-category is in [56].
to fix. to add: Barr; Hyland-Power-Cheng.
further reading:
1.FibBic, equipments, double categories, Street-Cauchy, yoneda, cosmoi, Veritythesis 2. Kelly-amiens, Power-lawvere, Power-bilimits 3. Street-stacks, Street2Dsheaf, Weber
Sections not covered: didn’t say much about Chapter 6, which I assume you’ve
heard a lot about already. For 8.6 see [54] and also [35]. For 9.5 the starting points
are [19] and [44] for the enriched stuff, and [42] for how to start to weaken things.
68
STEPHEN LACK
I didn’t talk about Chapter 10 except for the 2-nerves stuff; some relevant papers
here are [52, 53] for the 2-/bicategorical point of view; of course there is a huge
amount of stuff on stacks themselves. There is a “Giraud” notion of 2-topos in
[53] and an elementary version in a recent preprint of Weber; the two are quite
different. The latter involves a notion of subobject classifier where subobject is
taken to mean discrete fibration. For Chapter 11, see [50] and [51]. I didn’t really
talk about 12.5 and 12.7 which I imagine are very familiar around here. As for 12.6,
Yoneda structures go back to [55]; see also [49] for the related notion of cosmos.
Equipments were (and continue to be) studied by Wood in a series of papers with
various coauthors; see for example [61, 7, 6].
Acknowledgements. It is a pleasure to acknowledge support and encouragement
from a number of sources. I am grateful to the Institute for Mathematics and
its Applications, Minneapolis for hosting and supporting the workshop on higher
categories in 2004, and to John Baez and Peter May who organized the workshop
and who encouraged to submit this companion. The material here was based on
lectures I gave at the University of Chicago in 2006, at the invitation of Peter May
and Eugenia Cheng. I’m grateful to them for their hospitality, and the interest
that they and the topology/categories group at Chicago took in these lectures. I’m
particularly grateful to Mike Shulman, whose excellent TeXed notes of my lectures
were the basis for the companion.
References
[1] Jiřı́ Adámek and Jiřı́ Rosický. Locally presentable and accessible categories, volume 189 of
London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge,
1994.
[2] Jean Bénabou. Introduction to bicategories. In Reports of the Midwest Category Seminar,
pages 1–77. Springer, Berlin, 1967.
[3] Renato Betti, Aurelio Carboni, Ross Street, and Robert Walters. Variation through enrichment. J. Pure Appl. Algebra, 29(2):109–127, 1983.
[4] G. J. Bird, G. M. Kelly, A. J. Power, and R. H. Street. Flexible limits for 2-categories. J.
Pure Appl. Algebra, 61(1):1–27, 1989.
[5] R. Blackwell, G. M. Kelly, and A. J. Power. Two-dimensional monad theory. J. Pure Appl.
Algebra, 59(1):1–41, 1989.
[6] A. Carboni, G. M. Kelly, D. Verity, and R. J. Wood. A 2-categorical approach to change of
base and geometric morphisms. II. Theory Appl. Categ., 4:No. 5, 82–136 (electronic), 1998.
[7] A. Carboni, G. M. Kelly, and R. J. Wood. A 2-categorical approach to change of base and
geometric morphisms. I. Cahiers Topologie Géom. Différentielle Catég., 32(1):47–95, 1991.
[8] Denis-Charles Cisinski. La classe des morphismes de Dwyer n’est pas stable par retractes.
Cahiers Topologie Géom. Différentielle Catég., 40(3):227–231, 1999.
[9] Jean-Marc Cordier and Timothy Porter. Homotopy coherent category theory. Trans. Amer.
Math. Soc., 349(1):1–54, 1997.
[10] Charles Ehresmann. Catégories et structures. Dunod, Paris, 1965.
[11] Samuel Eilenberg and G. Max Kelly. Closed categories. In Proc. Conf. Categorical Algebra
(La Jolla, Calif., 1965), pages 421–562. Springer, New York, 1966.
[12] Peter Gabriel and Friedrich Ulmer. Lokal präsentierbare Kategorien. Springer-Verlag, Berlin,
1971.
[13] John W. Gray. Formal category theory: adjointness for 2-categories. Springer-Verlag, Berlin,
1974.
[14] André Joyal and Ross Street. Pullbacks equivalent to pseudopullbacks. Cahiers Topologie
Géom. Différentielle Catég., 34(2):153–156, 1993.
[15] André Joyal and Myles Tierney. Strong stacks and classifying spaces. In Category theory
(Como, 1990), volume 1488 of Lecture Notes in Math., pages 213–236. Springer, Berlin,
1991.
A 2-CATEGORIES COMPANION
69
[16] G. M. Kelly. Coherence theorems for lax algebras and for distributive laws. In Category
Seminar (Proc. Sem., Sydney, 1972/1973), pages 281–375. Lecture Notes in Math., Vol.
420. Springer, Berlin, 1974.
[17] G. M. Kelly. Doctrinal adjunction. In Category Seminar (Proc. Sem., Sydney, 1972/1973),
pages 257–280. Lecture Notes in Math., Vol. 420. Springer, Berlin, 1974.
[18] G. M. Kelly. A unified treatment of transfinite constructions for free algebras, free monoids,
colimits, associated sheaves, and so on. Bull. Austral. Math. Soc., 22(1):1–83, 1980.
[19] G. M. Kelly. Structures defined by finite limits in the enriched context. I. Cahiers Topologie
Géom. Différentielle, 23(1):3–42, 1982.
[20] G. M. Kelly. Elementary observations on 2-categorical limits. Bull. Austral. Math. Soc.,
39(2):301–317, 1989.
[21] G. M. Kelly. Basic concepts of enriched category theory. Repr. Theory Appl. Categ.,
(10):vi+137 pp. (electronic), 2005. Originally published as LMS Lecture Notes 64, 1982.
[22] G. M. Kelly and Stephen Lack. On property-like structures. Theory Appl. Categ., 3:No. 9,
213–250 (electronic), 1997.
[23] G. M. Kelly and A. J. Power. Adjunctions whose counits are coequalizers, and presentations
of finitary enriched monads. J. Pure Appl. Algebra, 89(1-2):163–179, 1993.
[24] G. M. Kelly and Ross Street. Review of the elements of 2-categories. In Category Seminar
(Proc. Sem., Sydney, 1972/1973), pages 75–103. Lecture Notes in Math., Vol. 420. Springer,
Berlin, 1974.
[25] Max Kelly, Anna Labella, Vincent Schmitt, and Ross Street. Categories enriched on two
sides. J. Pure Appl. Algebra, 168(1):53–98, 2002.
[26] Stephen Lack. On the monadicity of finitary monads. J. Pure Appl. Algebra, 140(1):65–73,
1999.
[27] Stephen Lack. A coherent approach to pseudomonads. Adv. Math., 152(2):179–202, 2000.
[28] Stephen Lack. Codescent objects and coherence. J. Pure Appl. Algebra, 175(1-3):223–241,
2002.
[29] Stephen Lack. A Quillen model structure for 2-categories. K-Theory, 26(2):171–205, 2002.
[30] Stephen Lack. A Quillen model structure for bicategories. K-Theory, 33(3):185–197, 2004.
[31] Stephen Lack. Limits for lax morphisms. Appl. Categ. Structures, 13(3):189–203, 2005.
[32] Stephen Lack. Homotopy-theoretic aspects of 2-monads. submitted, available as
arXiv.math.CT/0607646, 2006.
[33] Stephen Lack and Simona Paoli. 2-nerves of bicategories. K-Theory, 2006. to appear, available
as arXiv.math.CT/0607271.
[34] Stephen Lack and Ross Street. The formal theory of monads. II. J. Pure Appl. Algebra,
175(1-3):243–265, 2002.
[35] Aaron D. Lauda. Frobenius algebras and ambidextrous adjunctions. Theory Appl. Categ.,
16:No. 4, 84–122 (electronic), 2006.
[36] F. William Lawvere. Metric spaces, generalized logic, and closed categories. Rend. Sem. Mat.
Fis. Milano, 43:135–166 (1974), 1973. Reprinted as Repr. Theory Appl. Categ. 1:1–37, 2002.
[37] Saunders Mac Lane and Robert Paré. Coherence for bicategories and indexed categories. J.
Pure Appl. Algebra, 37(1):59–80, 1985.
[38] F. Marmolejo. Doctrines whose structure forms a fully faithful adjoint string. Theory Appl.
Categ., 3:No. 2, 24–44 (electronic), 1997.
[39] F. Marmolejo. Distributive laws for pseudomonads. Theory Appl. Categ., 5:No. 5, 91–147
(electronic), 1999.
[40] Paddy McCrudden. Opmonoidal monads. Theory Appl. Categ., 10:No. 19, 469–485 (electronic), 2002.
[41] I. Moerdijk. Monads on tensor categories. J. Pure Appl. Algebra, 168(2-3):189–208, 2002.
Category theory 1999 (Coimbra).
[42] A. J. Power. Coherence for bicategories with finite bilimits. I. In Categories in computer
science and logic (Boulder, CO, 1987), volume 92 of Contemp. Math., pages 341–347. Amer.
Math. Soc., Providence, RI, 1989.
[43] A. J. Power. A general coherence result. J. Pure Appl. Algebra, 57(2):165–173, 1989.
[44] John Power. Enriched Lawvere theories. Theory Appl. Categ., 6:83–93 (electronic), 1999.
[45] John Power and Edmund Robinson. A characterization of pie limits. Math. Proc. Cambridge
Philos. Soc., 110(1):33–47, 1991.
[46] Ross Street. The formal theory of monads. J. Pure Appl. Algebra, 2(2):149–168, 1972.
70
STEPHEN LACK
[47] Ross Street. Two constructions on lax functors. Cahiers Topologie Géom. Différentielle,
13:217–264, 1972.
[48] Ross Street. Limits indexed by category-valued 2-functors. J. Pure Appl. Algebra, 8(2):149–
181, 1976.
[49] Ross Street. Cosmoi of internal categories. Trans. Amer. Math. Soc., 258(2):271–318, 1980.
[50] Ross Street. Fibrations in bicategories. Cahiers Topologie Géom. Différentielle, 21(2):111–
160, 1980.
[51] Ross Street. Cauchy characterization of enriched categories. Rend. Sem. Mat. Fis. Milano,
51:217–233 (1983), 1981. Reprinted as Repr. Theory Appl. Categ. 4:1–16, 2004.
[52] Ross Street. Characterizations of bicategories of stacks. In Category theory (Gummersbach,
1981), volume 962 of Lecture Notes in Math., pages 282–291. Springer, Berlin, 1982.
[53] Ross Street. Two-dimensional sheaf theory. J. Pure Appl. Algebra, 23(3):251–270, 1982.
[54] Ross Street. Frobenius monads and pseudomonoids. J. Math. Phys., 45(10):3930–3948, 2004.
[55] Ross Street and Robert Walters. Yoneda structures on 2-categories. J. Algebra, 50(2):350–379,
1978.
[56] Zouhair Tamsamani. Sur des notions de n-catégorie et n-groupoı̈de non strictes via des ensembles multi-simpliciaux. K-Theory, 16(1):51–99, 1999.
[57] R. W. Thomason. Cat as a closed model category. Cahiers Topologie Géom. Différentielle,
21(3):305–324, 1980.
[58] Dominic Verity. Complicial sets: Characterising the simplicial nerves of strict ω-categories.
Mem. Amer. Math. Soc., to appear.
[59] R. F. C. Walters. Sheaves and Cauchy-complete categories. Cahiers Topologie Géom.
Différentielle, 22(3):283–286, 1981.
[60] R. F. C. Walters. Sheaves on sites as Cauchy-complete categories. J. Pure Appl. Algebra,
24(1):95–102, 1982.
[61] R. J. Wood. Abstract proarrows. I. Cahiers Topologie Géom. Différentielle, 23(3):279–290,
1982.
[62] K. Worytkiewicz, K. Hess, P.-E. Parent, and A. Tonks. A model structure à la Thomason on
2-cat. J. Pure Appl. Algebra, 208(1):205–236, 2006.
School of Computing and Mathematics, University of Western Sydney
E-mail address: s.lack@uws.edu.au